you need to train in nfloat4 which is horrible quality and doesn't learn properly in one trainer
you need to train in nfloat4 which is horrible quality and doesn't learn properly in one trainer
Munless you rent gpu 48gb gpu and train all in bfloat16
Monetrainer hasnt fixed fp8 yet
Mlocally not a good option almost 130gb
Fdid you follow this tutorial without skipping anyt step?
FYouTubeSECourses
Welcome to the Pillars of AI Application Libraries: How to Install Tutorial! 
Are you ready to embark on a journey of installing AI libraries and applications with ease? 
In this video, we'll guide you through the process of installing Python, Git, Visual Studio C++ Compile tools, and FFmpeg on your Windows 10 machine. 
We'll also show you ho...
Are you ready to embark on a journey of installing AI libraries and applications with ease? In this video, we'll guide you through the process of installing Python, Git, Visual Studio C++ Compile tools, and FFmpeg on your Windows 10 machine. We'll also show you ho...Fonetrainer has so many options 0 explanation lol
Fkohya way better
Fi started first training downloading files
Myeah i've done about 40 trainings now with onetrainer
Monly realized at the end when i tried all with bf16
Mthat all the nfloat4 ones lacked details
Fmy first test like this
F
F
Fi will start 8 trainings today
Ftext encodfer 2 is T5 XXL?
M@AIDarkView sd-scripts FLUX: te1 and te3
OneTrainer FLUX: te1 and te2
ComfyUI only supports the latter format for now.
In addition, there is a potential issue with the current implementation of ComfyUI that prevents it from loading T5XXL LoRA without CLIP-L.
OneTrainer FLUX: te1 and te2
ComfyUI only supports the latter format for now.
In addition, there is a potential issue with the current implementation of ComfyUI that prevents it from loading T5XXL LoRA without CLIP-L.
9/10/24, 1:29 AM
Fok this config using 16275 vram
Mone trainer loras introduced major issues to kohyas keys in comfy and forge
Mbreaking things
MFthat guy gave buch of kids to have so much power on his repo
Myeah quite bad atm.
either way kohya is better than onetrainer at this moment for flux
either way kohya is better than onetrainer at this moment for flux
Mkohya will improve even more with in a few days
Fstarting second training
Mare you training the full lora or just attn?
Mi wouldn't bother too much with dora, too much compatibility issues at the moment
Ffull lora
Fmn this interface slowness killing me
Mlogit normal is just like sigmoid i think
Myeah it was buggy for me on massed compute
Fdora doesnt work on swarmui?
Mi think only in comfy
Mas sometimes images are pink
Fshould go xformers or sdp here not sure
Fhe still didnt upgrade to torch 2.4
Mi used sdp, but xformers should work
Myeah one of the guys from onetrainer had to leave the project due to work restrictions
Ffloat8 fails lol
Myeah its a bitsandbytes error as he says but i guess he implemented it wrong
Msince it works in kohya fine
Fdmn interface on linux sucks
Msave your self a lifetime of headaches and stick with kohya for now
Fdid you try ema with flux?_
Myeah
Fye kohya config perfect atm
Mema improves
Mema with cpu
Fkk