🤝｜tech-support

[fair oxide]

Which cpu is best for stable diffusion amd ryzen 9 7950X3D or Intel core i9-14900k and how big difference is it

[inland zephyr]

Anyone using Automatic1111 and finding that SDXL checkpoints indefinitely stall at 50 or 95%, even with medvram-sdxl and xformer arguments?

[inland zephyr]

I'm reinstalling, since apparently this problem extends to non-XL checkpoints too

[fair oxide]

I tried this and everything worked smoothly under 5 steps

[dark owl]

Intel and nvidia have been the kings for over 50 years. Amd budgetware boys are sad when they get error for almost everything

[ornate elk]

Boath are good. Stable Diffusion dont rely on CPU, there wouldn't be any difference.

[fair oxide]

How do I import Civitai files into fooocus?

[tight vapor]

hi all! Anyone know what the hell is wrong with this one LOL

why she look like an ogre

[vestal viper]

hi, i'm really new in this stuff. tried to install SD and i've followed all the steps but i kept getting errors when trying to launch webui-user.bat. at first it showed error unable to use GPU or something (i'm using RX 6600) so i googled it and applied skip GPU check. now i got another error like runtime error couldn't clone stable diffusion. googled it and tbh i really have no idea what does anything mean like manually set path to python etc, i'm not tech savvy so can anyone explain it in easier words please?

[tight vapor]

mac or pc

[vestal viper]

pc - windows 10

[bleak ridge]

i'm trying to contribute (game pc doing nothing), but i'm receiving

[tight vapor]

ah idk about windows haha sorry!

[bleak ridge]

is this the right channel for worker questions?

[gilded river]

hello my DS is only taking up 5 gigs of vram, and any ideas?

[bleak ridge]

i guess 6GB GPU is not enought to contribute?

[tight vapor]

bruh like what the hell

[gilded river]

its wierd how its up to 7.3

guess that's cool

[gilded river]

Put white teeth on your prompt if you're already using restore faces

[tight vapor]

but

how about this one? The reference pic is a 9/10 gir

[gilded river]

lmao

[tight vapor]

and I get this

an ogre

[gilded river]

what model are you using

[tight vapor]

like wtf LOL

https://civitai.com/models/25694?modelVersionId=143906

this one

[gilded river]

I am not nearly skilled enough to help you

put more steps?

[tight vapor]

but why they so ugly HHAHAH oh my god its been bugging me all day

aight, ill try more! Do you know why it doesnt care about the reference pic at all?

[gilded river]

maybe get a textual invesion for negative prompts

[tight vapor]

aigth thanks bro

[gilded river]

haven't used reference pictures

[ornate elk]

Pls no nsfw content here

I've deleted the image

[tight vapor]

ahh sorrrry

my b

[ornate elk]

Np, for better quality you need to use hires fix

[tight vapor]

okay thanks!!

any tips on how to stick to consistent face? mine goes spastic.

and which model for most realism? :S

[robust jacinth]

Hey!
Every time I restart my RunPod session with Automatic1111 in Jupyter Notebook, I encounter errors from plugins indicating missing module dependencies. I'm unsure how to reinstall these dependencies effectively after each restart. Can anyone guide me on how to ensure these modules are available or how to reinstall them quickly when I start a new session?

[crimson leaf]

Im trying to use dreambooth and train my first lora, but it instandly ooms at the beginning, something im doing wrong?

RuntimeError: No executable batch size found, reached zero.

[karmic crown]

It worked only because you finally fixed Python. You would have the same issues as with A1111 if you tried to install that one first

if you go to your OS (C:) -> AI -> stabile-diffusion-webui
and run webui-user.bat
this time it should install everything and it should work

[fair oxide]

Thanks, it's dowloading rn. Do you have a list of reccomended prompts or prompt directions to follow?

What should I see after it is downloaded? @karmic crown

[fair oxide]

Okay it's downloaded and working fine but I have some memory issues apparantly

[spice warren]

Hi all, anyone knows what I do wrong here? I tried to use ControlNet to colorized my sketch but somehow it doesn't work out. The outcome never matched with my input sketch. I was playing with the numbers from Apply ControlNet and KSampler but still no luck yet.. and it requires a lot 200+ steps for KSampler. Please help..

[ornate elk]

whats your gpu ?

[fair oxide]

Geforce RTX Nivida 3050 laptop

[worthy portal]

Hi everyone
I encountered a problem in the installation of stable diffusion using auto1111 and no matter what I do, it is not solved
Please guide me if you know where this problem comes from
I will post the photo of anaconda as well

[ornate elk]

then you need to edit the webui-user.bat, at the line COMMANDLINE_ARGS=
you need to add: --xformers --medvram --no-half-vae
then save and relaunch.

[ornate elk]

delete the venv folder and try again.
if it still dont work you need to run
pip cache purge
then delete the venv folder again

[fair oxide]

Like this?

[worthy portal]

pip cache purge ?

[ornate elk]

yes

[ornate elk]

yea its a command oyu have to run in the anaconda terminal

[worthy portal]

ok I'll try

[fair oxide]

My pc restarts randomly when I open either foocus or SD

Yeah the screen just blinks, goes black, freezes for a bit and it cancels generating

[fair oxide]

Yep it's not working

WHY does this shit keep happening to me

[ornate elk]

How much RAM do you have ?

Also make sure your Laptop is plugged in

[fair oxide]

16 gb

[fair oxide]

Yes

https://media.discordapp.net/attachments/1188450266295570444/1188534772155301888/image.png?ex=659ae043&is=65886b43&hm=8b1535b6cb9b192b42907baf85923bd461e0bfb855540a31292dc6636b4eda5a&

My laptop is always charging

Nobody can beat this cursed laptop

[wind sun]

yo guys im with the new build and im trying to install the processor drivers and its failing the installation

someone ?

[worthy portal]

I implemented what you said, but I still encountered the same problem
.
I had previously installed stable diffusion on this system and it was running perfectly fine, but for some reason I deleted it and now when I want to install it again, I can't.

[ornate elk]

You may need to increase your pagefile and make sure to have enough free space on C:/

[ornate elk]

Maybe you need a BIOS update

[livid jetty]

I have a photo of my cat and i want to put him in a forest how do i do that?

[sturdy ginkgo]

think modifying a workflow for comfyUI might be the easiest .. use one that can generate a alpha mask of the cat (invert it) and have a forest background generated behind it.

https://civitai.com/models/237306/marduk191s-character-insertion-workflow

[livid jetty]

ok and how would i make and image look better, i generated a dog and i want the dog to look better

[ornate elk]

with upscaling

[burnt badger]

when using multiple loras do i need to put them in a certain order or does weighting take care of everything?

[ornate elk]

you need to play around with the lora strenghts

[ornate elk]

i would recommend you install SD with my install guide and not an all in one installer in powershell, if you need my guide let me know.

[tight vapor]

How do you upscale? Just send IMG2IMG and upscale by 2?

[nova hedge]

im getting this error when trying to install the automatic1111 webui

[ornate elk]

What's your GPU and which webui did you tried to install ?

[ornate elk]

There are 2 ways.
First is using hires fix in txt2img.
Or using SD upscale script in img2img

[nova hedge]

rtx 3070 and i git cloned from https://github.com/AUTOMATIC1111/stable-diffusion-webui.git

[ornate elk]

interesting, someonme above had the same error today

maybe something is broken

[nova hedge]

i might just have to keep trying cause sometimes it gets further into the download

[pliant tide]

Is there a huge trade off in performance, upgradability between a laptop like this and similarly priced/specc'd tower?

Gigabyte AORUS 17X C5AU Core i9 RTX 4090 17.3in QHD 240Hz Laptop

[ornate elk]

yes, mostly gpus in laptops are so called mobile gpus, they are downclocked gpus with lower performance so the laptop dont melt xD also sometimes they cut of vram.
3050 in pc = 8gb
3050 in laptop = 4gb
never buy a gaming laptop.
if you can afford a 4090 put it in a pc

[pliant tide]

Thank you.

[ornate elk]

np, imagine the battery time of a laptop with these specs. it wont last very long taking the point of portability. also a laptop with these specs need to be pluged in at any time to give the gpu enough power.
Airflow in Laptops is really bad. with these components i cant imagine the heat xD

[pliant tide]

Yeah, portability was the only real draw card, but that’s a good point on battery.

Guess I’m heading back to pcpartpicker then. Wish me luck 😬

[ornate elk]

yea, portability is the only point for a laptop. but not for gaming or high performance stuff. For Office or normal Users its perfect.
Good luck 🙂 whats your budget and Use case?

[pliant tide]

I’m purely building for an AI art workflow. Rest of my life is in Apple ecosystem. So looking to spend around 3-4K USD for a nicely future proofed system.

Doesn’t need to be pretty, just powerful.

[ornate elk]

ahh okay, then a Tower PC is the best choice. 4090 Asus Strix or MSI SuprimX are the best in performance. CPU either intel or AMD.
for PSU maybe 1200W BeQuiet, 64 or 128gb RAM, and a good case with a nice airflow (much space for Fans), Lian Li or NZXT have nice ones. Or of course BeQuiet (less rgb, but silent)

[burnt badger]

is there a way to get more controlnets? have 3 now, would like as many as possibe

[ornate elk]

You can go into the Settings under controlnet and increase the slider

[fair oxide]

How do I do that

Also after restarting everything there is another problem, I can't even press generate, it just reverts back immediately

Like it loads for a second and gone

^ this is for fooocus

[ornate elk]

launch sysdm.cpl from the Start menu search or run box (Win+R) and navigate to Advanced –> Settings –> Advanced –> Modify
Set it to "20000" for C:
And disable it for all other drives.
Then save and restart the PC.
Make sure to have 15-20gb free space on C drive.
Then restart your PC.

[pliant tide]

Perfect. Thanks. And yeh, the bequiet dark stuff is super slick. I built my last pc around that with tinted black glass case, with ‘dark knight’ theme. Thanks again and happy xmas from Australia.

[ornate elk]

Ouh that sounds nice! Bequiet stuff is great! Using their Power Supply's only. And their fans in my last PC.
Merry Christmas from Germany.

[fair oxide]

It just opens system propertes

[ornate elk]

Then click on the windows logo and type
Variables.
Click on the first entry it shows. Then follow my instructions from above

[fair oxide]

?

[ornate elk]

Hmm nope

@fair oxide
https://mcci.com/support/guides/how-to-change-the-windows-pagefile-size/

[ornate elk]

Lol wtf xD currently allocated 38gb

Do you have something running atm?

[fair oxide]

yea

what do i do

[ornate elk]

What do you have currently running?

[fair oxide]

I tried restarting my pc

I just opened SD / fooocus, and this shit happens

[ornate elk]

Dont open boath at once.
Make sure your PC is updated.
Then run
SFC /scannow
In a cmd as administrator.
That fixes windows errors.
I'm off for today

[burnt badger]

Is there a way to capture steps of the renderer as it progresses?

[fair oxide]

I didn't

Okay wait

How do I run SFC?

One sec

wtfwtf

[ornate elk]

You need to open the cmd as admin

To use that command

[fair oxide]

how do i do that

[ornate elk]

Type cmd in search. Then right click. Run as administrator

Also make sure you have the latest nvidia driver installed

If not then update it and delete the venv folder and relaunch the webui-user.bat

I'm off for now.

[livid jetty]

lets say i made a dog, it turned out great except the dog's eye was weird, is there a way for me to just fix the eye or do i need to start all over, when i use the img to img it gives me a completeley different image.

[dark owl]

@kindred tinsel I can help you in here

[kindred tinsel]

ah thank you thank you let me switch to pc then

This is what i am getting rn, i have the highest I figured would work for openpose normal and it doesn't detect anythinbg, also apologies for the image, not that sfw

Im not sure which model it uses to do this, and if i can change it to the v1.1 sd1.5 openpose that I already have

[dark owl]

ive never seen anyone use light mode before

it looks like it's working to me?

cept her clothes got in the way in that one

[kindred tinsel]

also when i change yolo models i get recursive errors in python, let me reproduce the error real quick

[dark owl]

a lot of the time it requires 2 controlnets, Open pose + Line art, or HED.

ill show you my node system

[kindred tinsel]

sorry bout that

I am using this, and get that error, does it mean that there is an error or incompatiblity between preprocessors, because it doesn't really give me much to go on :c

[dark owl]

there are 2 models Open pose models and DW open pose models and they don't mix

[kindred tinsel]

So should I put both like at the same time and plug both into the generation using that controlnet apply thingy?

[kindred tinsel]

Oh then that must be the issue since one is like dw-ss smth smth and the other is the yolo but with this torchscript thingy which I do not recognize

also do you know if I can edit the pose inside of comfy with this pack or do I need another one?

[dark owl]

this is your problem

use 1 or the other

not both

[kindred tinsel]

ahhhh oke oke oke I will, I think DW gives better results at first glance so I will try that one out

[dark owl]

yes DW is better

[kindred tinsel]

and can i edit the bone structure in comfy or should i do that in webui?

[dark owl]

Not sure i've never tried in comfy

I just use refference image

or a model pack from civit ai

https://civitai.com/models

can go here and filter "open pose" only

[kindred tinsel]

so when you do your animations, how do you manage the poses? do you change them a bit or like allow the model to do its own work?

Also I got it working now and finally I can generate, follow pose and upscale on the same workflow thanks for your help ^^

[dark owl]

Video2video, record yourself with phone camera and rip the poses from it.

[kindred tinsel]

oooooooh I see so like you get a pose for every keyframe and then do it, that makes more sense lmao

[dark owl]

[kindred tinsel]

thanks for the help ^^

[jaunty sphinx]

how do i solve this bruvs:

woa yooa posing like a woman

dats big brain

oh yeah how do i configure sd.next to run open vino

it says i shouldnt edit the webui.bat so should i edit something else or make a new file?

oh shit i figured it out using the cmd prompt

torch size still broken though

how do i fix the thing in the message txt.

why does the middle torch size cut in half

[dark owl]

maybe dont skip torch cuda compile is my only guess

https://developer.nvidia.com/cuda-toolkit

make sure this is up to date

[jaunty sphinx]

oh thanks

sd next open vino works but it makes my computer flash when the gpu is at 100% lol

how do i make it use only 90 percent lol

oi i dont have nvidia tho

[dark owl]

ohhh im no help then sorry Idk anything about AMD

[jaunty sphinx]

dam

[jaunty sphinx]

dam it works perfectly on sd next just fries the computer until i can't see anythang 😮

[topaz maple]

I have a problem with Adetailer:

When applying Adetailer for the face alongside XL models (for example RealVisXL v3.0, Turbo and Non-Turbo Version), the resulting facial skin texture tends to be excessively smooth, devoid of the natural imperfections and pores. This occurs even though Adetailer impressively enhances details in areas like the eyes and mouth. Consequently, the facial skin appears unnaturally airbrushed, lacking in realistic detail.

Any advice on how to fix this?

PS: I did not test if this issue also occurs with SD1.5 models.

Here are comparisons of images without (left) and with (right) Adetailer.

RealvisXL 3 Turbo:

https://i.imgur.com/PisIdpA.jpg

https://i.imgur.com/ET1JXpQ.jpg

RealvisXL 3 Non-Turbo:

https://i.imgur.com/tyZZdgI.jpg

https://i.imgur.com/j8uco63.jpg

My Adetailer settings:

https://i.imgur.com/sQsrmRD.png

[dark owl]

@topaz maple click use separate steps and try 30+

[topaz maple]

I will try!! Thanks! I will report back

[topaz maple]

I increased the steps. how does that look? is there an improvement?

[dark owl]

yes looks better

now it needs upscaler

[burnt badger]

Is there a way to download sample images for loras that you have installed? so the cards in the menu show you what they look like?

[topaz maple]

thank you for your advice. But I am a bit confused why I would need about 20-30 steps with a Turbo XL model for Adetailer, and for the primary generation the Turbo models are ok with 8 steps....

but increasing the steps definitly helps with the skin textire

[dark owl]

maybe it's the sampler? I find Euler A gives more realistic humans

[topaz maple]

it is weird, but I am glad I found a solution

thanks for helping out

[topaz maple]

I dont exactly how to do it, but I know that STability Matrix (https://github.com/LykosAI/StabilityMatrix) have done this in their checkpoint / Lora manager. I guess they access the metadata von civitAI models including example images via the CivitAI API: https://github.com/civitai/civitai/wiki/REST-API-Reference

[burnt badger]

sweet, i'll give it a try

lora folder is getting way too big and I am forgetting wtf i've been downloading

[dark owl]

there is an easy way with automatic1111, you can render an image then use that for the lora cover picture

[rotund lava]

civitai helper extension

[serene horizon]

amdpoor here
trying to get the directml thingy with olive to work but I get this weird error whenever I try to convert the model

    Traceback (most recent call last):
      File "G:\stable-diffusion-webui-directml\modules\call_queue.py", line 57, in f
        res = list(func(*args, **kwargs))
      File "G:\stable-diffusion-webui-directml\modules\call_queue.py", line 36, in f
        res = func(*args, **kwargs)
      File "G:\stable-diffusion-webui-directml\modules\ui.py", line 1759, in optimize
        return optimize_sd_from_ckpt(
      File "G:\stable-diffusion-webui-directml\modules\sd_olive_ui.py", line 67, in optimize_sd_from_ckpt
        unoptimized_dir, optimized_dir = ready(unoptimized_dir, optimized_dir)
      File "G:\stable-diffusion-webui-directml\modules\sd_olive_ui.py", line 35, in ready
        unload_model_weights(shared.sd_model)
      File "G:\stable-diffusion-webui-directml\modules\sd_models.py", line 881, in unload_model_weights
        send_model_to_cpu(sd_model or shared.sd_model)
      File "G:\stable-diffusion-webui-directml\modules\sd_models.py", line 632, in send_model_to_cpu
        if m.lowvram:
    AttributeError: 'NoneType' object has no attribute 'lowvram'

running with the flags --onnx --medvram --use-directml(this one is kinda scuffed cuz the official guide told to use --backend directml but apparently it doesn't work)

[lament tusk]

Hello, I have some issues to train dreambooth lora using diffusers

I am using this code but not working well when I change images

https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/SDXL_DreamBooth_LoRA_.ipynb

[worthy portal]

yes I need

[karmic crown]

civitai browser plus is better (or at least i like it more) 🤣

[karmic crown]

amd or nvidia?

[worthy portal]

nvidia T2000

[karmic crown]

#🤝｜tech-support message <- nvidia guide by CS1o

[worthy portal]

I do the same thing every time but I still run into problems

[worthy portal]

Could these errors occur due to the graphics card driver not being updated?

[karmic crown]

have you installed everything in a new folder?

also, please run git status and show me what you got

[karmic crown]

that is miniconda, use standarad installer, since automatic1111 use venv, you are now trying to use two diffrent virutal enviroments together

also, cs1o wrote to install in c:\AI since in some cases installation on desktop can cause a problem

you cant run git status in ai, you need to go inside stabile-diffusion-webui, and run it from there

[worthy portal]

Please see this as well, I am getting this error

[karmic crown]

i saw that

you can try to use newer pytorch

go to webui-user.bat and add this:
set TORCH_COMMAND=pip install torch==2.1.2 torchvision==0.16.2 --index-url https://download.pytorch.org/whl/cu118

then remove venv
and run webui-user.bat
it will create a new venv with newer version of torch

please let us know if that worked

i need to go now

work time starting

[worthy portal]

thanks verymuch

[junior raven]

hi, why when i try to DM images the bot is "thinking" for ever?

i mean, right click, apps and dm images, or this is something else?

[verbal olive]

do amd users still need to use linux to make use of their gpu properly with stable diffusion? particularly the 5xxx radeon series?

[fair oxide]

Are Macs better for SD and code?

[vestal viper]

please help, waht should i do?

[fair oxide]

AMD Ryzen 5 6600H with Radeon Graphics, 3301 Mhz, 6 Core(s),

My pc

I’m getting a new laptop soon anyway, it’s a Mac, would it be better?

[karmic crown]

which mac? its not the same if it is m1 or m3

Intel Macs = ☠️
M1 is slow as hell for SDXL, but it's not too bad for 1.5
M2 is better, but still not even close as good as nVidia

I have M1 MacBookPro and iMac 2019 (Intel + Radeon 580x)
M1 works faster with A1111

[ornate elk]

Yes

[ornate elk]

No

[fair oxide]

What laptop should I get

[ornate elk]

You need to git clone again

You had a connection error

[vestal viper]

so run cmd and git clone to the same folder?

[ornate elk]

You can also try to make a new folder to git clone in

[vestal viper]

aight gonna try, thanks

[karmic crown]

depends on your budget and your needs

Silicon Macs (m1, m2, m3...) work extremely fast with javascript (which includes apps made in Electron), much faster than any PC
on the other hand, Silicon Macs are slower for stabile diffusion than PC laptops with 3xxx Nvidia graphics

if you are buying a Mac, MacBookPro is always a better choice than Air with the same processor (my last 3 laptops were MacBook Pros)
Macs need less RAM than PCs, so MacBookPro with 16GB is usually the same as a PC with 32 GB for most things (PhotoShop, for example)

keep in mind that you can not add more memory to your Mac laptop later since memory is soldered to the motherboard; you need to order one with more memory on purchase
it is possible to change SSD, but I saw cases where a new SSD made a lot of problems - do not take anything with less than 500GB SSD

my recommendation for Mac is an M2 (or newer) MacBookPro with at least 16 GB of RAM (if you have enough money, take 32GB) and at least 500 GB of SSD (1TB in an ideal case)

for PC, anything with a newer processor and Nvidia 3xxxx with dedicated RAM should be OK; just do not forget that in some cases, the latest i5 works faster than the old i7, so check the numbers (larger is better) and not the naming (i5, i7, i9)

[ornate elk]

For Nvidia 8gb vram should be the starting point. The more the better

[karmic crown]

i agree with @ornate elk

[hoary sphinx]

Can I train using Mac? is there a way of training textual embeddings, hypernetworks, lora or fine-tunining? I have a M1 64Gb

[karmic crown]

it should work, but the question is how fast (slow)

[hoary sphinx]

it always gives errors when I try to create a Textual embedding in automatic 1111

[karmic crown]

what you have in webui-user.sh?

and what text (with version infos) you have on the bottom?

[hoary sphinx]

v1.6.0
#!/bin/bash
#########################################################

Uncomment and change the variables below to your need:#

#########################################################

Install directory without trailing slash

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

Name of the subdirectory

#clone_dir="stable-diffusion-webui"

Commandline arguments for webui.py, for example: export COMMANDLINE_ARGS="--medvram --opt-split-attention"

#export COMMANDLINE_ARGS=""

python3 executable

#python_cmd="python3"

git executable

#export GIT="git"

python3 venv without trailing slash (defaults to ${install_dir}/${clone_dir}/venv)

#venv_dir="venv"

script to launch to start the app

#export LAUNCH_SCRIPT="launch.py"

install command for torch

#export TORCH_COMMAND="pip install torch==1.12.1+cu113 --extra-index-url https://download.pytorch.org/whl/cu113"

Requirements file to use for stable-diffusion-webui

#export REQS_FILE="requirements_versions.txt"

Fixed git repos

#export K_DIFFUSION_PACKAGE=""
#export GFPGAN_PACKAGE=""

Fixed git commits

#export STABLE_DIFFUSION_COMMIT_HASH=""
#export CODEFORMER_COMMIT_HASH=""
#export BLIP_COMMIT_HASH=""

Uncomment to enable accelerated launch

#export ACCELERATE="True"

Uncomment to disable TCMalloc

#export NO_TCMALLOC="True"

###########################################

[karmic crown]

please add ``` text ``` at the beginning and the end, so it will be formated

but, ok, i see its not ok
where is your installaton located?
in which folder?

[hoary sphinx]

/Users/gustavomello/Local_Folder/07_MISC_TOPICS/stable-diffusion-webui

[karmic crown]

download my a1111 script from https://github.com/viking1304/a1111-setup/releases/tag/0.0.5
and run it like this (from the terminal)

sh a1111-setup.sh -f all -d /Users/gustavomello/Local_Folder/07_MISC_TOPICS/stable-diffusion-webui

it will

• update your a1111 to latest version (1.7.0)
• update torch to the latest version (2.1.2)
• add some fixes
• change your command line params (for better performance and compatibility)

or follow my step by step guide https://github.com/viking1304/a1111-setup/discussions/2
from existing install part

but i recommend script, since is faster

and MUCH easier 🙂

unlike on windows, on macs you will still launch ./webui.sh and it will call webui-user.sh

[hoary sphinx]

thank that is great actually!

[hoary sphinx]

have you tried training with Mac?

[fair oxide]

What do you recommend for nvida

[karmic crown]

not yet, but:

• 1.5 works at a decent speed
• SDXL works without any problems (but is slow)
• upscale works

I honestly didn't play too much, since I spend a lot of time on speed testing, parameter optimization and making a script and tutorial

everything else also should work

i would be really glad if you can test it and report the results, so i can add that as a remark at the end of guide

[karmic crown]

i would not dare to recommend any PC laptop, since i have not used PCs for years

[ornate elk]

a 3060 with 12gb, a 4070 12gb or 4070ti if you want to game good on it too.

a 3060 is still good for all games in FullHD, 40xx series have more power

[karmic crown]

@ornate elk do not foget, that he asked about laptop with nvidia

[ornate elk]

ohhh

right

[karmic crown]

thats why i said i would not dare to recommand anything, if it was just a graphics card i would wrote same as you 🙂

[ornate elk]

i dont recommend any Laptop for gaming or SD.
but if a laptop is needed then any nvidia gpu with 8-12gb vram will do it

[chrome rampart]

@ornate elk merry christmas and thks for all your help this year

[karmic crown]

the only PC laptop that i would ever recommend to anyone is this one https://rog.asus.com/laptops/rog-zephyrus/rog-zephyrus-g14-2023-series/
with a remark that is a based only on description and that i haven't seen it live

[ornate elk]

hey thanks, and merry christmas 🙂

[karmic crown]

i had i7 asus gaming laptop many years ago (before ROG), and i liked that one a lot (it had 17" screen) 🙂

merry christmas to you, mine is in 2 weeks 🤣

[fair oxide]

Thanks I will check it out

[karmic crown]

but please keep in mind that my recommendation is based only on the description of that laptop and my personal experience with an old Asus gaming laptop I had many years ago

[hoary sphinx]

dude, thanks a lot! training embeddings with mac is faster than with my 4090! amazing

[karmic crown]

how?!

please let me know about the timings later

[fair oxide]

@ornate elk

I was told to delete the venv, I tried many times, closed and restarted everything

It keeps returning

[karmic crown]

this looks good

how do you run this?

are you running webui-user.bat?

[fair oxide]

Except when I try to generate a prompt with a model that I followed instructions give, copy pasted the extention; it doesn't work

In fact no models work

The generation comes out garbled

[fair oxide]

Yes

[karmic crown]

ok

let it finish

and show me the bottom of your user screen

with version numbers

[karmic crown]

still good

23.0.1 is good enough, you do not need to update that

[fair oxide]

Can send a model and example prompt to test out

Maybe it's just the specific models I'm using that are fucking with me

[karmic crown]

wait

let it finish

and show me what you have here

i am intrested in the version infor at the bottom

and make a screenshot when you finish creating image so we can see if you did something wrong

[ornate elk]

You should try Dreamshaper v8

Then show as the result

[main axle]

Hello, I am trying to create a sdxl lora using diffusers

I am using this script to train a model, but it doesn't work well.

Does anyone can help me?

[sonic ferry]

ComfyUI, what do parenthesis do, how do you use them? Is there a limit?

[kindred portal]

I downloaded SDXL Turbo and stable diffusion 2.1, the only one that works is stable diffusion 1.4. I have 4gb vram and 16gb ram

I checked and I had enough ram left while rendering

[kindred portal]

Hey, that ui looks interesting/useful. What application is this?

[ornate elk]

Thats automatic1111 webui
The most used one here and in general

You should also use auto1111 for better optimisation with your GPU.
But sdxl won't work with 4gb vram or really really slow

[teal cedar]

new error popped up recently, didn't have it till around today

[ornate elk]

You installed automatic1111 the wrong way. (Inside a Onedrive Folder)
That causes problems. You also uses a different installer than normaly used.

[teal cedar]

The installer was a must for me to be able to utilize my GPU. It worked fine until today.

[ornate elk]

What's your GPU ?

[teal cedar]

AMD Radeon RX 6600

[ornate elk]

Ah okay. Normaly you could use the Directml version of auto1111

[sturdy ginkgo]

install to a folder like C:\SDWebUI\ instead and it should work fine ... OneDrive etc loves to disallow access randomly (security purposes as cloud storage rehashes the files making direct access only available to windows itself without redownloading)

[autumn inlet]

What is the break point of whn you should/shouldn't use med VRAM?

[bleak shoal]

i keep having this Warning for Canny contolnet. but im sure everything works fine. does anyone has any ideas why the warning?

[kindred tinsel]

Heeyyyy people, merry christmas first, now onto the question.

Is there a comfyUI node, or way, to save images with a structured name. I don't want to save the images with the normal ComfyUI(seed).....

I want

ComfyUI-Result-(number)-(date)

I have some nodes that allow me to concatenate strings, but I am having a hard time wrapping my head around how to manage the number, I produce in batches of four and the images are outputted in tandem each iteration so like how can I send each of them to save with a different number?

[ornate elk]

6gb vram and below is --medvram

[orchid adder]

Are the inpainting versions of ckpts better for ADetailer?

[autumn inlet]

oh... so I shouldn't have been using it with my 1080ti then?

I got a 4070ti for xmas so I guess I don't need it now for sure 😛

[ornate elk]

yes, you should only use --xformers

[autumn inlet]

what about no half vae?

[ornate elk]

yes that helps to for sdxl and anime vaes

[autumn inlet]

sweet, I was already noticing a massive speed increase a project on my 1080ti took 18m40s, I re-ran it with my new GPU and it took 4m20s

[ornate elk]

18 to 4 minutes wow

[sturdy ginkgo]

could be you are using a 1.1 or 1.5 controlnet model with SDXL or vice versa .. it cannot see it on the filename that is why it says the warning

[karmic crown]

I just remembered you said you have 64GB. Then you could probably make some things even faster

open your webui-user.sh
and add # at the start of your current command line args
and then add this one below

export COMMANDLINE_ARGS="--skip-torch-cuda-test --opt-sub-quad-attention --upcast-sampling --no-half-vae --use-cpu interrogate --disable-model-loading-ram-optimization"

[karmic crown]

Batch Size 4, Steps: 10, Sampler: Euler a, Size: 1024x1024, Model: hephaistosNextgenxlLCM_v20, Version: v1.7.0

I went from 11 min. 8.3 sec. to 8 min. 30.4 sec. a few days ago

and now:
Time taken: 7 min. 51.6 sec.

I ditched both medvram and no-half on my M1 MacBookPro with 16GB, and I have no errors 🙈

[wind sun]

yo do someone know how i can fix fingers

[karmic crown]

i used to use some text inversions, but i do not use that anymore

[wind sun]

is that for me

or

[karmic crown]

yes

i am now using just this negative prompt

deformed, bad anatomy, disfigured, poorly drawn face, mutated, extra limb, ugly, poorly drawn hands, missing limb, floating limbs, disconnected limbs, disconnected head, malformed hands, long neck, mutated hands and fingers, bad hands, missing fingers, cropped, worst quality, low quality, mutation, poorly drawn, huge calf, bad hands, fused hand, missing hand, disappearing arms, disappearing thigh, disappearing calf, disappearing legs, missing fingers, fused fingers, abnormal eye proportion, Abnormal hands, abnormal legs, abnormal feet, abnormal fingers

[wind sun]

i will try now

thanks

[karmic crown]

i used to use badhandv4, ng_deepnegative_v1_75t, FastNegativeV2 and similar

[wind sun]

well now i have to find a chechpoint for full body shots cuz i have one of the best ones but its only from waist to head

[karmic crown]

photorealistic or anime? 1.5 or sdxl?

[wind sun]

i need it to be as real as possible

[karmic crown]

i love those two

https://civitai.com/models/200061?modelVersionId=256258 SDXL
https://civitai.com/models/84728?modelVersionId=90072 1.5

[wind sun]

well is it only civitai for checkpoints

[karmic crown]

there are some others, but i mostly use civitia

[wind sun]

alr

[wind sun]

i will try them later \

[karmic crown]

i also use https://www.liblib.art/ but if you cant read chinese or do not have chinese friends, you will probably be lost there 🙂

[bleak shoal]

Thank you, i have to check it again to make sure but i don't think im using sdxl.

[kindred tinsel]

Hello one question.

How different are, performance wise,

1. Nvidia 4060 TI 16gb vram
2. Nvidia 3090 ROG 24gb vram
3. Nvidia 4090 Ti 24gb vram

Is veam more important that’s series like is the 3090 going to perform closely to the 4090 or would it be better to buy the 4060 although it has less vram?

[sturdy ginkgo]

3090+ it is less about VRAM and more about TensorCores

4060 out performs 4090 Ti on that

[kindred tinsel]

So like I’m confused which one is better for this kind of workflow? My card rn struggles with up scaling and basic generation so I want to have one that I can push hard on work; animation, etc

[sturdy ginkgo]

buy one your wallet allows

usually it is not the gpu but a different component causing issues ... dying HDD/SSD.. RAM not running optimal

[kindred tinsel]

Really? I can fit the 4060 but I’m scared the 12 gb vram isn’t enough. I’m running with 8gb and most of the time I get memory alloc errors. I was thinking of keeping 16 gb ram and the 3090 with 24 gb

I mean I’m sure the 3090 will be slower than the 4060 or 4070 but more vram sounds good

[safe inlet]

in the world we live in now the 3090 is faster than the 4060 it seems

https://www.tomshardware.com/pc-components/gpus/stable-diffusion-benchmarks

[ornate elk]

there is no 4090 ti,
in SD, vram is all that matters. a 3090 is much better and faster than a 4060 or 4070
tensor cores dont get used in SD, only when installing nvidias tensorRT extension.

[wind sun]

Any idea how can i remove clothes with sd

[wind sun]

I mean im with 4070 rn and there is almost 2 seconds for generations

[ornate elk]

Simply prompt what's under it xD

[wind sun]

Wym

Im stupid, you know that

I need more steps

i created that AI model but i want it without the bra

[ornate elk]

For example if the character wears a jacket. Prompt
Wearing T-shirt.

[wind sun]

look at your DM's

[kindred tinsel]

So they have a random card up in their store? :0

[kindred tinsel]

So probably a 3090 would be good since it has tons of memory? Also for training does that not require cuda cores or smth like that or only vram?

[ornate elk]

Yep a 3090 or 4090 are the best for training cause of the vram amount.

For example sdxl requires a lot of vram to train
They use cuda cores but not tensor cores

[kindred tinsel]

Okay I see, well I found a 4090 for 2299 and a 3090 for 1100 both at 24 gb lmao I think I’ll push for the 3090 then

[fair oxide]

reasonable. also consider psu upgrade.

[kindred tinsel]

Yeah I’m thinking about it too cus after taxes prices grow a lot. And also I need to get more storage ram and ssd so like

I need to replace every single thing in this computer

[fair oxide]

SSD storage is a must for optimum SD performance. maybe around 500gb. Ram could be fair with 32gb.

[kindred tinsel]

I have 128 rn but my entire setup runs through a HDD. I was thinking of keeping the models there and moving executables and the like over to the ssd

Ram is expensive tho :( I’ve got 16 (2*8)

[fair oxide]

128gb ssd?

[fair oxide]

16gb ram might not be all that performing with sdxl btw. recommended is 32gb but people even go up to 64gb which is not strictly necessary.

[kindred tinsel]

Yeah but it’s full just with windows installation

[fair oxide]

right, SSD drives are costly compared to HDD but you should aim for at least 500gb SSD.

[kindred tinsel]

I don’t really work with sdxl I’ve never tried it since by GPU struggles with sides of 1024 * 1024 but I wanted to get into it once I get capable hardware

My first idea for a model was something about agriculture, realistic landscapes for agriculture etc. Easy to categorize based on crop, growth rate, season, terrain

[fair oxide]

reason for using SDXL is far better prompt coherence than SD 1.x versions. once you get to use SDXL its hard to ignore.

[kindred tinsel]

Hey is it okay if I dm you so I don’t bog down the chat?

[fair oxide]

its fine to talk here.

much better for our learning is anyone else shares vital info too

[kindred tinsel]

Aight I found one ssd says “NV2 500Gb M.2 PCle NVMe” for 57 dollars is that good ?

[fair oxide]

yes the nvme ones are good faster than other variants.

but for actual pricing you have to check locally

[kindred tinsel]

Nah but like 24 more and you get 1 TB

[fair oxide]

not a bad deal if its in your budget range

[kindred tinsel]

Looks like this

That’s supposed to go in the motherboard right ? Just underneath the the cpu or smth

[fair oxide]

yah it goes into supported mobo like a ram stick

place it in and screw it in .. easy to install

[bright hill]

I've got a weird bug that's happening to me

[fair oxide]

https://www.backblaze.com/blog/nvme-vs-m-2-drives

[bright hill]

Whenever I boot up Auto11, Lora files aren't loading in

When I look at the webui.bat, it says "model.safetensors: 48%"

It's taking forever to load

This hasn't happened before

[fair oxide]

nvme at 500gb isnt too pricey btw

[kindred tinsel]

I’m thinking of revamping storage like this

500gb or 1 tb ssd for applications, windows startup and basically just localized central installations of comfy and webui.
3tb HDD models, datasets, image results

[kindred tinsel]

Yeah but I heard m.2 is like faster and rn I’m using a 128 m.2 and it is quite fast

[fair oxide]

if you want to leverage on SD speed models need to be in SSD too

[kindred tinsel]

So you are telling me I need to move 280 gb of models into the ssd?

There literally won’t be any space left, besides how many read and writes operations does sd do on the models once loaded in vram? Wouldn’t them being in the HDD affect only loading times?

[fair oxide]

the difference might not be too high btw, and since you are having to store 128gb worth of models i think what you said about storing them in HDD could be a better alternative

but im not entirely sure about running SD on SSD and read models from HDD

[kindred tinsel]

Yeah I’ve got a ton of them and mostly what stops me from running generations is the lack of speed for generation with AMD

[ornate elk]

Its okay to store models on hdd

It won't affect generation time

[fair oxide]

ahh thats nice

[ornate elk]

I store a lot on HDD and using symlink to just load them normal in SD

[kindred tinsel]

Thank youuuu :))) so that’s one thing less to worry about

[fair oxide]

you just opened up a solution for me @ornate elk

[kindred tinsel]

Using what?

Is that for Nvidia or smth never heard of that

[ornate elk]

Nope, its an OS thing.
My webui is installed on an SSD.
SD thinks all my models are in models/Stable-diffusion but I have linked them to a folder on my HDD.
SD then just loads them normaly as if they were in the folder on the SSD.

You can do that with any file or folder

[kindred tinsel]

So how does that work? You just reference them inside the models directory inside of like idk webui settings?

Like how you add extra models paths for comfy?

[ornate elk]

Yes

But without changing any config path

I just let SD think everything is in the right place

Symlink stands for Symbolic link.

[kindred tinsel]

Oh and if I do get an ssd would you mind helping me out to set up that later?

[ornate elk]

Sure 🙂

[kindred tinsel]

I’m probably going to buy these two things before I move back to uni. I’ll die a little bit more every day either way with how slow Amd is

Once I have money I’ll buy the 3090 and finally be able to sit down and milk the power of AI

[ornate elk]

With 24gb vram you won't get any problems with ai stuff.
Your then also be able to use large language models and voice ai stuff

Pretty fun stuff

[kindred tinsel]

Yeah I was thinking of doing smth like a little chatbot trained on calculus books and proofs just as a proof of concept. Like a model that regurgitates proofs already done but that in any case learnt something and I can call my own haha

[safe inlet]

i dabbled a little with local LLM to get prompts

[ornate elk]

Also a 2tb formfactor 2.5 ssd is also good and not to expensive

[kindred tinsel]

A what and what now?

[fair oxide]

btw @kindred tinsel here is a suggestion about running SD from nvme card and storing models elsewhere. instead of buying hdd for storing you could buy regular SSD but not nvme to cut down on price a lot and avoid some of the HDD related bottleneck or just regular perfomance

[kindred tinsel]

Ah I seeee well yeah that could work but I’m not sure which to buy then. I’m aware that m.2 is the best and have always looked for it I don’t know any other models sorry.

My professor didn’t cover this in our intro to cs

[ornate elk]

For the OS always use an M2

[fair oxide]

i would suggest buying a 500gb nvme for running windows and other apps, but for storage purpose use something like basic ssd cards

[ornate elk]

Samsung and Western Digital make the best

Kingston is okay too

[kindred tinsel]

Hmmmmm but like I’m not sure tho. Is there an actual difference aside from speed ? Are they more reliable than a simple HDD?

[sturdy ginkgo]

lifespan of HDD is longer but SSD are faster and quieter

[fair oxide]

HDD are also prone to defect where SSD are far more robust

[ornate elk]

Yes. HDDs can fail and get corrupt. If you shake your PC while its running, you will break your HDD and you can lose Data for example.

SSDs are not impactable by physical shaking or moving

[kindred tinsel]

Well with six different fans noise is not that much of an issue tbh.

So then which model is the one ssd that’s chest and simple like an upgraded HDD.

[sturdy ginkgo]

lost 4 SSD's due to power outages :/

[kindred tinsel]

I will try and get the terabyte ssd for system. HOLD UP ? I’m having like daily outages due to seasonal rains being a little….. gone for power generation in my country; are those issues common?

[sturdy ginkgo]

Have system critical stuff on HDD \

[fair oxide]

@kindred tinsel this is just a suggestion btw if your budget allows, but in overall about regular SDD vs HDD for storage, SSD are better in performance and less prone to damage

[ornate elk]

Power outtage while the PC is running can make Data and Windows system files corrupt.

[kindred tinsel]

Dang I’m doubting what to do now :(…. I’ve moved my computer a ton these past years, driving it around for uni and nothing has happened to the HDD. But now I’m scared having more of them maybe idk I don’t wanna lose my sd stuff

[kindred tinsel]

Really? I’ve never had that even when literal lightning fell close to my house: my system rebooted and everything was back to normal :0

[ornate elk]

If the PC is Off, it won't do anything if the shaking isn't to hard

[kindred tinsel]

I mean I always lay the pc down flat, turned off; power supply off and wrapped with a seatbelt

[safe inlet]

power spike can kill SSD, mechanical failure kills HDD generally

[fair oxide]

its advisable to use a decent PSU for that like corsair as an example

[kindred tinsel]

Yeah but I mean this can happen any time, anywhere, and to both so the issue here is how you care about your computer

[fair oxide]

and fyi those corsiar PSU are mostly designed to die than to let anything happen to your PC in case of abnormal spikes

[bright hill]

Hey Volv,do you remember me? I'm the guy that apparently never had a Lora folder

I found it

[ornate elk]

I bought a power socket with 60.000amp over voltage protection

[bright hill]

Do you know what VTRoot is?

[fair oxide]

fine lines of products here.. very reliable in most case scenarios even power outages / spikes https://www.corsair.com/us/en/c/psu

[kindred tinsel]

No idea lmao maybe @ornate elk

But glad you found the Lora folder :))

[bright hill]

Basically, it's a file folder that's in C:\VTRoot

Except VTRoot doesn't show up normally

[kindred tinsel]

Also @fair oxide How can I check the psu information inside my pc. I don’t remember max watt

[bright hill]

Anyway, I feel like I'm exploring some spooky part of the computre, since I have no idea why this stuff is here

[fair oxide]

use the CPU-Z freeware tool

[kindred tinsel]

Okay I’ll give that try

[ornate elk]

Seems like a folder created by third party antivirus Programm

[bright hill]

That's what I'm reading

Comodo

[ornate elk]

Do you have that?

[bright hill]

Yes

[ornate elk]

OK then thats the cause

Would recommend Windows Defender over any Third Party antivirus software

For the normal users thats enough

[fair oxide]

btw, i dont see much option for PSU info through software, you have to open the casing and physically check for it

and most psu should have a bold lettering on it to show the wattage

[kindred tinsel]

Aaaaah okey okey I don’t wanna do that tho 😨 I’m scared I might kill it

I’ll just have my brother check on it when he instwllls the other parts

[sturdy ginkgo]

good modern PSU's have a LCD screen showing every little detail

most midrange pre-build systems have 650W or 750W

[fair oxide]

anyone updated their a1111 to 1.7.0 and seeing this attached warning message in the console?

[fresh geyser]

Recently re-installed A1111 and somehow dont have the ability to drag new inpainted image over the old image i was inpainting to replace it? any idea why?

[indigo ivy]

how do i fix this

[fair oxide]

Give prompt

[sturdy ginkgo]

move or reinstall to a folder on your hdd that is not part of OneDrive (C:\SDWebUI\ for example)

[fair oxide]

@ornate elk

Stops generating and this happens

[sturdy ginkgo]

another instance is already using the Cuda API .. check if any other python processes are active in taskmanager, if so force close/ or reboot pc

[indigo ivy]

im moving it to downloads and its notbudging at all, this normal?

[rose crystal]

they removed it i think

[main axle]

Does anyone used this one with sdxl?
https://github.com/huggingface/diffusers/blob/main/examples/community/regional_prompting_stable_diffusion.py

[karmic crown]

this is your problem
Launching Web UI with arguments: --lowvram

you must have a lot more params for your configuration, the ones you had before and just to replace medvram with lowvram, not to replace all of them with just lowvram

i need to work, cu guys later

[indigo ivy]

help

[ornate elk]

You have to add --use-directml
To the webui-user.bat

At the line COMMANDLINE_ARGS=

[indigo ivy]

its working now thx

[ornate elk]

No problem 🙂

[fair oxide]

@ornate elk any chance you might be running a1111 v1.7.0 ?

[ornate elk]

Yes with nvidia and amd

[fair oxide]

i just got it updated and im getting a warning message about ...

[ornate elk]

Would recommend deleting the venv folder

[fair oxide]

this warning message that im getting #🤝｜tech-support message

and i checked the line 2967 and its already is what it asking me to change

but other than the warning message i dont notice any errors

[ornate elk]

Ok. try delete the venv and let it rebuild

[fair oxide]

oooh... can i ignore that mesage tho?

since i dont see any glitches with generating images

[ornate elk]

You can but idk what it will cause

Better to refresh the venv

Maybe its gone then

[fair oxide]

ok

thanks, i will see if there is any other possible workaround

this warning message is new, so if i dont get any other solutions i'll delete the venv and update it

[ornate elk]

Make sure to update your extensions too

Also do you had the tensor rt extension installed? That could be something from it

[fair oxide]

i installed tensor rt before but removed it later, that was several weeks ago, and that warning message popped up few hours ago after git pull

[ornate elk]

That could be a rest of that extension

That now got triggered due some changes in auto1111

Stuff installed by extensions in the venv won't get deletet when you delete an extension

Thats why its recommended to delete the venv from time to time.
After big updates. Or after deleting heavy extensions

[fair oxide]

hmm that sounds logical, i'll delete the venv folder and update.

also didnt realize until you said it now that it may be a good practice to refresh the venv folder in between big updates

[ornate elk]

Yea thats how I do it after big version changes.

[chrome rampart]

morning all. i am getting new gpu YAYYYYYY. Is it needed/or smart to ones installed the new card to remove my VENV folder and regenerate?

[ornate elk]

Yes

Better to recreate the venv.
If not it could think your still having less vram from the old gpu

[chrome rampart]

thks CS already think so. also cloning my HDD>SD to SSD drive so should see huge time improvement yay

[ornate elk]

Yea, but moving the webui can breake the auto update function

That can be fixed with a few commands

[chrome rampart]

thks for the extra tip will sure check and might get back to you then if ok

[fair oxide]

I just put the bare minimum

Subject
Medium
Style
Artist
Website
Resolution
Additional details
Color
Lighting

So my pc is just too bad for it

[ornate elk]

What's your GPU again?

And what's in your webui-user.bat?

[fair oxide]

silicon Mac’s are slower for stable diffusion
How much slower

[fair oxide]

AMD Ryzen 5 6600H with Radeon Graphics, 3301 Mhz, 6 Core(s),

[ornate elk]

Wait you have no dedicated GPU?

And what's in your webui-user.bat?

[fair oxide]

thanks for the suggestion, big help. my webui is running smoothly again with no warning messages. 🙂

[ornate elk]

So it works now?

[fair oxide]

Yes I think the problem comes when I put more than 3 words in the positive prompt

[ornate elk]

Can you check in task manager under performance what's your GPU?

You need some additional commands

[fair oxide]

Images still come out garbled in models

[vocal burrow]

How so ? Any example ?

[dense sandal]

I was trying the img2vid using svd and got this error. strange because it worked about a week ago.

I updated everything and still got the error

[ornate elk]

Okay then edit your webui-user.bat
By right clicking it.
Where you added --lowvram
You also add: --xformers --no-half-vae
Then save and relaunch the webui-user.bat

[clever nebula]

hello i am getting this error all of a sudden:

RuntimeError: mat1 and mat2 must have the same dtype

[ornate elk]

Thats mostly a resolution error

When using loras or models that are not for the same base version

[clever nebula]

im not using any loras and im using the recommended resolution in the models description

are these args okay?

[ornate elk]

What's your GPU?

[clever nebula]

4070 ti

[ornate elk]

Then remove --precision full
And add --no-half-vae

[clever nebula]

working thankyou sir

[idle monolith]

hello, i'm wondering what kind of hardware configuration i need for using stable diffusion.
If you got some recommandation or link to some configs, or just point me to the right direction.

[ornate elk]

Any nvidia gpu with a minimum of 8gb vram is good.
More vram is always better.
Best budget/vram is a 3060 12gb.
Good GPUs are also 4060 or 4070.
For CPU any modern CPU with 6 cores is enough. AMD has very good and cheap ones.
32gb RAM is recommended.

[idle monolith]

i was wondering if something like a minining rig with lots of card is a good idea or not at all ?

[fair oxide]

Okay currently works
Why are Macs worse for SD but good at video editing

[vocal burrow]

Video editing doesn't need tensor cores

[idle monolith]

because they are specialized. And if you do big big project for video editing, mac is too expensive

[vocal burrow]

Different hardware for different usage

[ornate elk]

Macs dont have dedicated vram like a GPU. They don't have Cuda Cores

[ornate elk]

You only can run a generation on one GPU.
So you can't combine the power of the cards

[idle monolith]

oh, good to know !

[ornate elk]

But you can generate multiple images at the same time with multiple gpus

Each running their own instance of sd

[idle monolith]

ok, so it's still faster with more gpus, but it's because it's just parallel work

[ornate elk]

Yep, depends on the gpus, but mining GPUs should have much vram

But a radeon7 isn't as fast as a 1080

[idle monolith]

so, do you have a dream config link somewhere, with no budget limit ? 😄

[ornate elk]

xD nope different countrys have different stores and prices. Dont have a general config rn.
But if you have no limit a 4090 is the best you can get.

Or an professional nvidia GPU

A100 xD

[dense sandal]

why svd is not working anymore? it used to work....

[fair oxide]

Dell G15 or Zeph G14

[dense sandal]

Error occurred when executing KSampler:

unsupported operand type(s) for *=: 'int' and 'NoneType'

please help, I updated every custom node

[chrome rampart]

question. Is there way to check were my python actually is installed?

[ornate elk]

Open up a cmd.
Run
where --python

[chrome rampart]

thks cs doubting if i better do new install python and SD on new SSD or clone my entire older HDD which includes both and lots more. thks again

[ornate elk]

Would recommend to install python in its default path on C.
Git cloning the webui is easily done in seconds.
Then just move the models folder over and the outputs of course.
Maybe extensions too.

[chrome rampart]

thks

one more. i think i would need to remove my "old" python right? can you tell me how?

[ornate elk]

Yea you uninstall it like any other software

Trough windows system control

Or in the new ui. Settings > Apps

[chrome rampart]

a easy lol thks

[ornate elk]

@grave zephyr Hey, follow my Install guide for automatic1111:
#🤝｜tech-support message

[reef jewel]

just downloaded my first extension(roop). i dont see it anywhere have i done something wrong?

i placed it inside C:\Users\myuser\auto1111\extensions

[chrome rampart]

@ornate elk did step you said and now works on SSD and YAYYY already lots faster. coming days new GPU and then cant blink ahahha kjoking> thks again

[ornate elk]

Nice! 🙂

[fallow tusk]

any thoughts on why this huge variance in generation times?

it seems I have to minimize the browser because it's using up too much GPU by itself so it starts lagging?
I have this issue on AUTOMATIC1111 but not on comfyUI, which is odd to me, but imo proves that it is an optimisation issue not hardware limitations

Sdxl turbo doesn't feel very turbo when I'm waiting 10 seconds per generation...

They were significantly faster on ComfyUI but idk how to do infill with it

[sturdy ginkgo]

if you mean inpaint: https://civitai.com/models/128556?modelVersionId=256787 try this and follow the youtube vids 🙂

[proper pike]

Does anyone know what type of settings to use to train a custom model in 1.5? I cannot figure out what settings to use on a 3090

[ornate elk]

What's your GPU and what's inside your webui-user.bat?

[ornate elk]

@nimble valve
for a 6700xt use these:
--use-directml --update-all-extensions --medvram --opt-sub-quad-attention --opt-split-attention --upcast-sampling

and dont use olive. its very experimental and dont support much

[nimble valve]

its a 6700S laptop cpu
is that similar to an xt?

and okay

ill send the dxdiag one sec

[ornate elk]

@nimble valve ah well the 6700s is the gpu

then yes use my commands

and if you still get out of memory at 768x768 then change --medvram to lowvram

[warped ginkgo]

So, what's this on about?
I've just upgraded to 1.7.0 of auto111's, uninstalled and re-installed python, and deleted the xformers folder in the users file, to GIT the one they talk about in the wiki

[ornate elk]

your torch version is to old, do you have dreambooth installed ?

[warped ginkgo]

No, I don't

[fallow tusk]

Rtx 3070 8+8 GB ram
I do have to check, it should be default settings since I haven't manually edited anything

[ornate elk]

okay then you need to edit the webui-user.bat and at the line COMMANDLINE_ARGS=
you add: --xformers --medvram-sdxl --no-half-vae
then save and relaunch

[nimble valve]

got it, will give it a shot now

does it matter which hires i use?

and any step/cfg recommendations?

[fallow tusk]

Ok thanksss
I'll have to check another day, already got fed up since I was struggling so much with my current prompt 😛

[warped ginkgo]

am i okay to just pip install torch==2.0.1

[ornate elk]

first try without hires fix.
then if you try with hires fix try the following settings:
512x768, 30 steps,
hires fix:
hires steps: 10
upscale by: 2
Denois: 0.5
Upscaler: Esrgan4x+

[ornate elk]

can you show me the full cmd output after launching the webui-user.bat?

[warped ginkgo]

i did and got the red

oh sure

[nimble valve]

oki, will try this

hmm got this error

[ornate elk]

it maybe wont work, but we can fix that with an extension called tiled diffusion

[ornate elk]

oh that means your webui isnt on the latest version

[nimble valve]

how do i update it

[ornate elk]

add git pull
to an empty white line before the webui.bat

[ornate elk]

whats your gpu?

also the cmd looks normal now

[warped ginkgo]

remind me how to check, if you don't mind?

[ornate elk]

task manager, performance, gpu

also check the version at the bottom of your webui in browser

[nimble valve]

like this?

[warped ginkgo]

Geforce RTX 4070 16gb

[ornate elk]

okay then remove --no-half from your webui-user.bat, its not needed for you and only slows down

[ornate elk]

versions look updated 👍

[warped ginkgo]

danke

[ornate elk]

bitte 😄

[nimble valve]

is it normal for the terminal to just display lines of "already up to date" and then close itself

[ornate elk]

it shouldnt close itself

[nimble valve]

thats what it did for me

after filling the screen with the quotes i sent

it just closed and did nothing

[warped ginkgo]

Is this important?

[nimble valve]

now it doesn't launch at all

when i run the bat file it just crashes instantly

[ornate elk]

nope, but you can download styles if you want, or create your own

[nimble valve]

running this instantly crashes it now

no error message no nothing

[ornate elk]

whats your folder called?
stable-diffusion-webui or
stable-diffusion-webui-directml ?

[nimble valve]

the first one

do i have to change the name?

[ornate elk]

then you installed the webui for nvidia gpus

AMD needs a different one

[warped ginkgo]

Aight. Then as per tradition, have Gordon Ramsay punching a cow.

[nimble valve]

welp i guess i'll move my models lora's vaes, and embeds and reinstall

[ornate elk]

Before following the guide, open up a cmd and type
pip cache purge
to remove the nvidia stuff.

Here is a quick Guide to install Automatic1111 Directml AMD Webui (Stable Diffusion)

You need to install Git 64bit: https://git-scm.com/download/win
and Python 3.10.11, 64bit: (any python above or 3.10.6) No 3.11, 3.12 https://www.python.org/downloads/release/python-31011/
And check "add python to path" when installing Python.

1. Make a new Folder on your drive (not on Desktop, downloads, documents, Programms, Onedrive) and name it Ai for example:
   C:\Ai\
2. You go into the folder you created in this case Ai, then click in the File Explorer bar (not searchbar) and type cmd then press enter.
3. Then you copy and paste this command:
   git clone https://github.com/lshqqytiger/stable-diffusion-webui-directml && cd stable-diffusion-webui-directml && git submodule init && git submodule update
4. Press enter and after its done you can close the cmd
5. Edit the webui-user.bat (right click), At the line COMMANDLINE_ARGS= You add: --use-directml --medvram --opt-sub-quad-attention --opt-split-attention-v1 --no-half-vae --upcast-sampling
   After that save and launch the webui-user.bat inside the stable-diffusion-webui folder.
   After the Installation that can take a while. Youll get an URL http://127.0.0.1:7860
   Thats the webui you open in Browser.

[nimble valve]

it launched, but there's no tab for VAEs

[ornate elk]

You have to manually add that.
Settings - User Interface - Quicksettings - there add SD_vae

Then apply and reload

[nimble valve]

oh alr

oh thanks i'll try now

[nimble valve]

can i leave the git pull line in there btw so it stays up to date?

also uh
where did clip size go

do i add that myself i assume

as expected, classic runtime error with hiresfix

without hires it renders perfectly

[ornate elk]

yes

[nimble valve]

wdym yes

well anyways
thanks
this works mostly

"RuntimeError: Could not allocate tensor with 655360 bytes. There is not enough GPU video memory available!"

uh what

[ornate elk]

every time you get this error you need to restart the webui to clear the vram

[nimble valve]

[ornate elk]

if not you will get that error again sooner

[nimble valve]

so refresh and it will fix itself?

well i do plan on getting an nvidia soon anyways
amd is pretty annoying for ai stuff

[solid ridge]

Why are my images do bad And don’t look like what I typed in

[nimble valve]

some models have recommendations for more specific/simple prompts

can also use a lora to help depending on what it is

[solid ridge]

What is a Lora

[nimble valve]

not sure how to explain it but basically if u wanna make a specific character u can find a lora of them and it helps a lot

"To put it in simple terms, the LoRA training model makes it easier to train Stable Diffusion on different concepts, such as characters or a specific style. These trained models then can be exported and used by others in their own generations."

[solid ridge]

Where do i find liess

[nimble valve]

civitai.com

has a lot of models reference images, etc...

[solid ridge]

Where do I put the downloaded one

[nimble valve]

stable-diffusion-webui-directml\models\Lora

if its a lora that is

if its a model, it goes in Stable-Diffusion instead of Lora

[solid ridge]

I got a .pt file, what is this

[nimble valve]

pytorch file iirc

usually stuff is .safetensors or .pt iirc

[solid ridge]

So is it s lora or model

[nimble valve]

it should tell u which it is when you download on civitai

it will say lora/model on the post

[solid ridge]

After I put a model into stable diffusion folder then how do I use it in the stable diffusion program

[keen wharf]

why are my images so bad pls help somebody

[nimble valve]

refresh it

and itll show up

well refresh the models tab

should be a button beside it

a blue button to the right

[keen wharf]

i have modell in my ai but it just creates random images

scout

!!!

pls

[solid ridge]

Ok but now even with the model when I type Taylor swift it’s not Taylor swift

[keen wharf]

help

me

[nimble valve]

uh

depends on what ur trying to make

diff models better for diff things

also gotta play with the settings

[nimble valve]

https://civitai.com/models/86858/taylor-swift

try this taylor swift lora

put it in the folder where i showed u and refresh the lora tab, then apply it

for the prompts add something like 1girl, taylorSwift, celebrity, etc...

to use lora's just click the lora and itll add it in the prompt

[keen wharf]

should i add anything else to the prompt aswell

[nimble valve]

depends on what ur trying to make

[solid ridge]

Why are you being so Transphobic?

[nimble valve]

well first of all what do u wanna make

[solid ridge]

Well it’s nsfw Type shi

[keen wharf]

i wanna make taylor swift as a donkey

[solid ridge]

Wow that’s very unethical

[keen wharf]

https://tenor.com/view/taylor-swift-cheers-taylor-swift-raising-glass-gif-42796408281448888

[nimble valve]

that lora i posted isnt nsfw only

they just put an nsfw example

you can add (NSFW) in negative prompts to get SFW only pictures

[fair oxide]

trying to get my comfyui workflow neat & tidy but these primitive nodes keep resetting after refresh even when pinned

any fix for this?

[ornate elk]

you need the tiled diffsuion extension

the tiled vae option can help against that error

[nimble valve]

how do u install it

[ornate elk]

go on extensions tab, click on Available, click on load from, search for tiled.
install tiled diffusion ( multidiffusion)

then restart the webui

[nimble valve]

the second one?

[ornate elk]

yes

[nimble valve]

it crashes less thanks

cant do 2.0 but now 1.5 works well

[ornate elk]

make sure to only enable tiled vae and not tiled diffusion

if you get gray squares, then set Endocer to 768 or 1024 and Decoder to 128

[nimble valve]

how do u do that

[nimble valve]

uh i cant find that

[ornate elk]

have you restartet after installing the extension?

[nimble valve]

i did

[ornate elk]

it should be an option at txt2img

at the bottom left side

[nimble valve]

oh i see it

this looks good?

tiled diffusion is disabled

[ornate elk]

set encoder to 768 or 1024, if 768 wont work 1024 should work always, and decoder to 128

[nimble valve]

forgot to say thx for the help
I can upscale up to 1.8 now without crashing

[ornate elk]

nice, no problem 🙂

[ornate elk]

check the cmd window

how big is the model file?

[ornate elk]

you still need to edit your webui-user.bat and add:
--xformers --no-half-vae

[ornate elk]

this is not a model. thats a lora

loras go into models/lora folder

[fair oxide]

Like this?

[ornate elk]

make a space between them

--lowvram --xformers --no-half-vae

[nimble valve]

hmm now encorder tile side 768 is causing the webui to crash after the photo finishes

[ornate elk]

make sure to close every other programm, dont run wallpaper engine in the backround, close all other browser tabs.

[tired estuary]

hey guys, i generated an image with a person (face and hair) i like and now i want to reuse the same face and hair in a different setting for example in a park or in the city with different clothes is there an easy way to do that? i tried pasting the image into imgtoimg and write a prompt and generate but it becomes a different person completely
help would be much appreciated 🙂
(sorry for copypasting from general)

[neat lagoon]

Hello!
Unfortunately I have got a little problem with my python version.
My command window tells me that the python version 3.11.5 is activated.
How can I change that to 3.10.6? I have not got python 3.11.5 installed anymore. The only version installed on my pc is 3.10.6 rn.
For context: I am currently at minute 37 of "ARPS2 Tutorial #1: Training Your Model" from the corridor Digital website tutorial.

[ornate elk]

hey, what programm are you using?

[neat lagoon]

I am using "EveryDream2trainer" (If you mean that) ... I am not really into this tbh

I am just trying to make some animated videos for a project.

[nimble valve]

alr

[ornate elk]

hmm okay, i didnt used that tool, but uninstalling python and installing 3.10 one with checked "addd python to path" should work

if the tool has a venv folder, delete it

[neat lagoon]

well... I tried to see the Checkbox but it was not there. I could for some reason not click any box...

[ornate elk]

if you run the setup of python again, click on modify, then next and then on "add python to system variables"

[neat lagoon]

By any chance... Just hypothetically: Is there any chance, anyone on this server could try and make a video (in anime style) to my selfmade music and my instructions? haha
Would be way easier for me (Yeah I am being a bit egoistical lol) and I actually do not understand anything of the stuff that's being said in the tutorial. I just understood, that I could ask in this Discord, if I had any questions, which I really appreciate.

[neat lagoon]

oh on modify. Yeah wait lemme try that one

[neat lagoon]

that worked! Now I just gotta put it into my folder (I think) and then check again... Thank you very much, kind sir! 🙂

[sturdy ginkgo]

i am trying to get comfyUI to use python 3.12 with the torch nightly builds but when starting it it says the safetensors module are not installed (when checking they are installed) .. any idea what i have wrong? ... changed the path inside the bat file to use the python312 executable instead of the embedded one in the package.

python 3.12 is in system variables path

[burnt badger]

anyone have any idea why SD is creating two copies of every image created? Wasn't doing this a couple days ago.

[young goblet]

Using A1111 WebUI? Check to see if one is .png and the other is .jpg. If so it's due to either the filesize or resolution. By default it saves a compressed .jpg copy of large images. Thresholds for that can be adjusted in settings or it can be turned off completely.

[young goblet]

Using the portable version of Comfy or manual installed?

[burnt badger]

I think you are correct, and I've just deleted the pngs since I thought they were copies. Do the jpgs contain the meta data?

[sturdy ginkgo]

from the provided zip, so the portable

[young goblet]

no, only the PNG versions will have meta data.

[young goblet]

Did you use the unstable nightly pytorch zip from here? https://github.com/comfyanonymous/ComfyUI/releases

[sturdy ginkgo]

ah no.. i used the one directly from the main page .. gonna try that one now

[young goblet]

yeah those portable versions are kinda funky if you want to modify them since they don't just use a python venv. I prefer manual install so I can easily update the deps manually. that nightly pytorch one should get you up and running though.

[acoustic drum]

How did fix Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0

[ornate elk]

open the nvidia control pannel, on the left click on 3D Settings, then on the right select the top dropdown and choose "Nvidia Highperformance Device"

[acoustic drum]

Do i restart the webui after that because the error keeps popping up

[ornate elk]

yes

also what webui did you installed?

[acoustic drum]

Its the text generation

By oobabooga or smthing like that

[ornate elk]

ahh the oobabooga

[acoustic drum]

Still nothing

[acoustic drum]

Not sire if 3d settings have any effect on text generation

[ornate elk]

What's your GPU ?
Maybe the model is too big

[acoustic drum]

Rtx4070 in lenovo slim5 laptop

[spark forge]

guys i'm experiencing the weirdest thing, i've been using stable diffusion (specifically a1111) for a very long time and my results have never been... this bad. the results suddenly look super bad, washed out and noisy, almost like i'm not using any vae. i have tested and it affects all of my models, including img2img. i have tried reinstalling multiple times and deleting the venv folder but the results are the same if not worse im using all of the same prompts as i have in the past and they were giving me the most beautiful and detailed results, i have no idea why it suddenly looks like this sampling steps do almost nothing, i can barely see any difference changing it from 20 up to 150. any idea why its suddenly like this?

[acoustic drum]

Model is dolphin-2.6-mixtral-8x7b

[acoustic drum]

By cognitivecomputations

[sturdy ginkgo]

check that any update to A1111 didn't reset your bat file to the default ... can cause lots of changes in the outputs, same for the configs

[ornate elk]

Seems like a missing vae

[ornate elk]

How big is the file?

[acoustic drum]

Like 20x5gb

So 100 maybe

[ornate elk]

Yea that won't work on any gaming GPU xD

[sturdy ginkgo]

yea 100GB ... that is for A100 GPU's etc

[acoustic drum]

Well the model loaded successfully

[ornate elk]

Can you show the file on your hard drive ?

[acoustic drum]

Well its all in the models file in the text-generetaion-webui

[ornate elk]

All these 5gb files ?

[acoustic drum]

Yep

But im not sure if the model beeing to big would explain the error „Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0“

[ornate elk]

Yea that won't work with your GPU

[acoustic drum]

Ok do you know a smaller model to try out

[ornate elk]

That errors can appear if the model dont fit in the gpu vram and gets loaded into ram too. And thats not usable

[sturdy ginkgo]

instead of oogabooga i would say use StudioLM instead ... has a build in ML model download and check if your system can run said model

[acoustic drum]

Well when i load the model it asks me to assign a certain amount of gpu memmory

[acoustic drum]

Tried it but the model that should be uncensored was censored even if it was the same model

[spark forge]

i am using vae

[acoustic drum]

For example, memory assigned 0

[acoustic drum]

Wait for it to load

[ornate elk]

Which VAE ?

It doesn't get applied correctly

Check the cmd

[spark forge]

in this case orangemix.vae.pt

[ornate elk]

@acoustic drum
Here you find models you can use:
https://gpt4all.io/index.html

[ornate elk]

What's your webui version?

[acoustic drum]

Then this error appears, tried to allocate 112 mib

[spark forge]

https://github.com/AUTOMATIC1111/stable-diffusion-webui/releases/tag/v1.7.0 reinstalled multiple times today

[ornate elk]

Hopefully not by downloading it as zip.

[acoustic drum]

I gib in the gpu, of which 7.2 are allocated by pytorch

[acoustic drum]

8

[spark forge]

no the issue started suddenly about 2-3 weeks ago, without any reinstalls

[ornate elk]

@acoustic drum your GPU is to weak for that model. Use a smaller one from the link

[acoustic drum]

But Then when i up the gpu memory in the webui the model loads just fine

[acoustic drum]

What link

[acoustic drum]

This link

[ornate elk]

#🤝｜tech-support message

If they won't work with oobooga install GPT4all

[acoustic drum]

But that dont look like a model at all

Its a whole instalation

[ornate elk]

Scroll down, there is a model list

[acoustic drum]

To where exactly

[ornate elk]

Model Explorer

[ornate elk]

Have you tried an other VAE ?

[acoustic drum]

Do you mean the dataset on huggingface

[ornate elk]

Lol its missing in your screenshot

On my phone its there under model explorer

[sturdy ginkgo]

shift+win+S is better for making screenshots 😄

[acoustic drum]

One of these maybe

[acoustic drum]

Nah dont have discord on the laptop yet

[ornate elk]

Nope its not there

[sturdy ginkgo]

is a windows feature

[ornate elk]

But hes on phone lol how should he send them xD

[sturdy ginkgo]

web version of discord 🙂

[ornate elk]

True

[acoustic drum]

Nvm i see them in different browser

So which one would you recommend

[sturdy ginkgo]

Still would use LMStudio and not get these CLI issues at all from oogabooga or gpt4all

[ornate elk]

Firefox should show them too but maybe an extension is blocking

[acoustic drum]

But the model there is weird, its censored when it should be uncensored

[acoustic drum]

Ill fix that later

[sturdy ginkgo]

were you using the swarm feature .. some are censored even if the model is uncensored

[acoustic drum]

Wait imma install it again

[mint seal]

I don't recommend that anyone use LMStudio.

[acoustic drum]

Which one should i choose?

[sturdy ginkgo]

They are legally obligated to state that for most countries

[ornate elk]

The smallest. But I doubt that it will work on a Laptop

[acoustic drum]

Ill try the q5 for now

Seemed to work before even if censored

[ornate elk]

Okay

[mint seal]

Regardless of any supposed obligations, these are incredibly intrusive and disturbing terms. I certainly haven't seen any such terms on Oobabooga or Sillytavern.

[ornate elk]

Or gpt4all

Is lmstudio even open source ?

[mint seal]

Nah, it's closed source.

[ornate elk]

Ah rip

[mint seal]

Btw, thanks for always fighting the good fight in helping people here CS1o. You probably don't get thanked a lot, heh.

[acoustic drum]

Then how else am i supposed to run the 100gb dolphin model

[acoustic drum]

True words

[ornate elk]

Thank you! Always here to help if I can 🙂

[mint seal]

I can't speak for the size, but I run LLMs with a combination of Oobabooga text generation webui, and SillyTavern.

They're not the most user friendly to set up, but they work very well and are quite compatible with most models out there.

[acoustic drum]

Well i get the error „Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0“ , model loaded fine

[ornate elk]

How much vram do you have ?

I tried oobabooga once with my 8gb GTX1080 and it failed

[sturdy ginkgo]

read through the whole thing .. and they are talking about that the program can and will update things like Python etc without notifying you in order to keep the program to work.. and any suspected modifications to LMStudio will be investigated for harmful conduct (aka versions spread with injected malware etc not by LMStudio)

[acoustic drum]

Here we go, theoretically the same model. 32gb in lmstudio. I use the advised threatening promt

Error, failed to load model, no such file directory

[sturdy ginkgo]

check the download manager if it succeeded the download especially these large ones as they download in parts

[acoustic drum]

It doesnt see the model under my models

[mint seal]

VRAM? 10gb. That all just depends on your model tbf. My system ram is 64gb though. I usually only use GGUFs, as I mainly run 13b models at 8bit quantization for best quality.

[mint seal]

The language used in those paragraphs that I screenshot isn't worded so specifically. They can do whatever they want with it update-wise, and look in on your usage without legal issue. That's directly how it reads.

[acoustic drum]

Ill just download the model again

[acoustic drum]

So any idea what the error means? Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0

[mint seal]

Have you looked up the error to see if anyone else has had it, and how they solved it? I'm not immediately certain as to any potential solutions for it myself, though the error looks familiar to my eyes.

I don't even know what you're running 😅

[acoustic drum]

Yes just google this exact line but damn i don understand anything in code

[acoustic drum]

https://www.google.com/url?sa=t&rct=j&q=&esrc=s&source=web&cd=&ved=2ahUKEwjs-c7Esq6DAxVR8QIHHUaCAcMQFnoECBQQAQ&url=https%3A%2F%2Fstackoverflow.com%2Fquestions%2F66091226%2Fruntimeerror-expected-all-tensors-to-be-on-the-same-device-but-found-at-least&usg=AOvVaw0Y5n0DaR8eN-xSjMVp2oet&opi=89978449

[acoustic drum]

Dolphin mistral8x7b in oobabooga on rtx4070 in lenobo laptop

[acoustic drum]

This error pops up when i try to chat with the ai and it respods with nothing

[mint seal]

Yeah, I'm no coder myself. The best way I've found for such things is to google and look for people talking about the same problem on places like Reddit, where there's convos on how other people have approached such things. That and searching such specific error codes in the Discord search function for the server.

I'll take a search myself.

[mint seal]

This may be beyond my knowledge then. From what I understand, Mixtral takes a special installation process that I am completely unfamiliar with.

[acoustic drum]

You can install it with anaconda, but i just did the old school internet download

[mint seal]

Mixtral is a model. It's not a AI front/back end, as far as I understand.

Meaning you'd need something like Oobabooga, LMStudio, llama.cpp, etc. to use it.

[acoustic drum]

Yes im using oobabooga

But this error just keeps me from doing anything with mixtral dolphin

[mint seal]

Mixtral is still relatively new, and some LLM AI implements may not be entirely compatible, or may require some special setup to work properly with it. If it were me trying to use it, I'd try to find a guide that involves a specific installation process for using the model with something like llama.cpp or ooba.

[sturdy ginkgo]

Might have to force the GPU in advanced settings of the webUI or in the config or Bat file .. try adding -h when loading the bat in the command line for possible config options

[acoustic drum]

Didnt understand that

[mint seal]

Here's a reputable guide that I've found for using Mixtral: https://www.reddit.com/r/LocalLLaMA/comments/18jxehq/guide_to_run_mixtral_correctly_i_see_a_lot_of/

[acoustic drum]

Lets try lmstudio now first

With the simple uncensored break into a car question

Im surprised it worked, the last time i asked it it refused to answer

But its quite slow and the cpu percentage jumps to as high as 600%

[acoustic drum]

How do i fix that

Gpu offload?

[sturdy ginkgo]

yea..seems it is using your CPU instead of GPU .. check settings if Use CPU is checked, if it autodetected to that than something with your drivers might be finky