@Dr. Furkan Gözükara Why 1536x1536 pixel images in the datasets? Training above 1024 is not a good i
@Dr. Furkan Gözükara Why 1536x1536 pixel images in the datasets? Training above 1024 is not a good idea right?!
Ait doesnt matter because of bucketing, they will get resized automatically to fit
Ajust wanted to ask if some people on here use 4090s for dreambooth training in kohya and their experience? I noticed that on runpod, a 4090 is a lot slower than a 3090ti and im trying to figure out if this is a issue on runpod or with a 4090 overall. Ive seen some issues posted on the kohya github about 4090s being slow, but im not sure if this is still the case
BHello Furkan, I have watched this video a couple of times and followed the instructions.
I have installed torch using the following command:
pip3 install torch==2.1.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
and transformers using the following command:
pip install xformers==0.0.23
I also edit webui-user.bat to use xformers:
set COMMANDLINE_ARGS=--xformers
With pip install xformers==0.0.23, I get the following:
Installing collected packages: torch, xformers
Attempting uninstall: torch
Found existing installation: torch 2.1.0+cu118
Uninstalling torch-2.1.0+cu118:
Successfully uninstalled torch-2.1.0+cu118
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
torchvision 0.16.0+cu118 requires torch==2.1.0+cu118, but you have torch 2.1.1 which is incompatible.
torchaudio 2.1.0+cu118 requires torch==2.1.0+cu118, but you have torch 2.1.1 which is incompatible.
Successfully installed torch-2.1.1 xformers-0.0.23
And when I execute webiu-user.bat, I get the following:
RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check
Is there something I am overlooking?
Thank you
I have installed torch using the following command:
pip3 install torch==2.1.0 torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118
and transformers using the following command:
pip install xformers==0.0.23
I also edit webui-user.bat to use xformers:
set COMMANDLINE_ARGS=--xformers
With pip install xformers==0.0.23, I get the following:
Installing collected packages: torch, xformers
Attempting uninstall: torch
Found existing installation: torch 2.1.0+cu118
Uninstalling torch-2.1.0+cu118:
Successfully uninstalled torch-2.1.0+cu118
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
torchvision 0.16.0+cu118 requires torch==2.1.0+cu118, but you have torch 2.1.1 which is incompatible.
torchaudio 2.1.0+cu118 requires torch==2.1.0+cu118, but you have torch 2.1.1 which is incompatible.
Successfully installed torch-2.1.1 xformers-0.0.23
And when I execute webiu-user.bat, I get the following:
RuntimeError: Torch is not able to use GPU; add --skip-torch-cuda-test to COMMANDLINE_ARGS variable to disable this check
Is there something I am overlooking?
Thank you
Ais it possible to train dreambooth on kohya with 24gig of vram? no matter what i try, it keeps running oom with 8bitadam
A@Dr. Furkan Gözükara Hey there - it seems there is an issue with download the runpod installer for your caption scripts
A
AAare you using the runpod kohya template by any chance? i had never noticed until watching the good dr's video that if you don't kill auto1111, 25-30% of the vram is getting sucked up before you start
Ai typically just reduce the batch size until it works, or throw on xformers, but i'm going to retry a few of them without cheating since apparently its bad juju
A@Dr. Furkan Gözükara aha, it's in the zip file at the end. maybe remove the 3 apparent links in that article when you have a sec 
Dcan anyone, help me fix this??
maybe this is the reason, my gpu isnt using full potential, while running the training!!
maybe this is the reason, my gpu isnt using full potential, while running the training!!
LLooks at potential to me, what are you looking at? You should update Kohya to 22.3 though
DOh, i see, yeah, let me do it as well. potential, means, when i run training, it doesnt take load on gpu it should take....usually, when the temperature gets higher, when the gpu is totally using it's power, but when i run training, with kohya, it only gets half of the temperature, compared to it usually takes.
Bsince today there's 22.4
Loohhhh What changes?
What changes?DOh, i can just edit the bat file and add autoupdate, it should update automatically, everytime it runs?
BGitHub
2023/12/28 (v22.4.0)
Fixed to work tools/convert_diffusers20_original_sd.py. Thanks to Disty0! PR #1016
The issues in multi-GPU training are fixed. Thanks to Isotr0py! PR #989 and #1000
--ddp_gr...
Fixed to work tools/convert_diffusers20_original_sd.py. Thanks to Disty0! PR #1016
The issues in multi-GPU training are fixed. Thanks to Isotr0py! PR #989 and #1000
--ddp_gr...
LOhhh seems useless unless you're running more than one GPU
BPagedAdamW added as optimizer and for the rest mostly fixes, but also some new options for sampling
LOk now im interested, whats pagedadamw
Lok yeah im about to do a training it looks interesting
LWWhat's the command to update Kohya again I forgot?
Lnevermind just a git pull
LLet's hope nothing is broken now
LIt broke blip ughhhh
LThe only thing that broke, it seems, was Blip captioning, but as I recall, I was using some type of rollback version of something. I think it was transformers. Honestly most of these captions tools are lousy and if you are using a smaller data-set manual captioning is best
Mdoes this mean we can do BF16 training now?
RCan I train a lora model for sdxl-turbo?
Fyes not but i still put if someone tries or uses
Fwhat is your it / s ?
Fhello. are you installing on activated venv?
Fwhich app_
Fyes i am doing that all the time. uses around 17 gb
Ftrue please download attachments. they got updated
Ftrue now in those zips
Fno it is not an issue. also you can upgrade kohya to latest
Fit is not currently at you
Fsadly not . it is gpu type limitation
Fi dont know yet sadly
Mthank you to use both gpus should we add "-ddp_gradient_as_bucket_view" to additional parameters?
MThe issues in multi-GPU training are fixed. Thanks to Isotr0py! PR #989 and #1000
--ddp_gradient_as_bucket_view and --ddp_bucket_viewoptions are added to sdxl_train.py. Please specify these options for multi-GPU training.
--ddp_gradient_as_bucket_view and --ddp_bucket_viewoptions are added to sdxl_train.py. Please specify these options for multi-GPU training.
Mhow do we specify these options in kohya?
Fyes i think
Fextra parameters
XCan I do multi-gpu training in Koyha if the graphics cards are different models/different Vram? (example, RTX 3080 and RTX 3090). Sorry to ask before testing, it's that I need a new power supply if I want to use both cards.
Ait is around 2s/it, a 3090ti has 1.12s/it on the same settings
Athe only time when the 4090 is faster, is if i increase the batch size to 4, its around 30% faster than the 3090ti then. But its still a bit weird
LDoes anyone know the Syntax for the regional prompter well. I'm getting better with prompts but if I have two Character Loras what's the best way to set up the prompt?
![Digital [Starburst]](https://cdn.discordapp.com/avatars/220961909447458817/8ecbbb3132fc6e44312f3d6d2ae3df69.webp?size=40)
DI'm getting [virtualMemoryBuffer.cpp::nvinfer1::StdVirtualMemoryBufferImpl::resizePhysical::140] Error Code 2: OutofMemory when attempting to generate the TensorRT Default Engine
DIt also says the Engine is invalid and I need to verify the engine was built correctly
