nvm, downloaded another model from kohya and replaced that model with eva02, it works
nvm, downloaded another model from kohya and replaced that model with eva02, it works
X
EHas anyone received this issue while trying to traing flux with Kohya?
Used the recent single click installer V15 from SECourses Patreon.
Used the recent single click installer V15 from SECourses Patreon.
EI know it is probably something to do with accelerate
Fwhich base model
Fyou given?
Ethe base model form the Blacklabs repo on hugging face
Efrom*
Ffile name and size please
FYouTubeSECourses
Many people having issues due to inaccurate Accelerate setup. This is how to setup to make it work properly
1-click installers and best configs for Kohya FLUX training :
A Big complete tutorial very soon hopefully : https://www.patreon.com/posts/kohya-flux-lora-110293257
1-click installers and best configs for Kohya FLUX training :
A Big complete tutorial very soon hopefully : https://www.patreon.com/posts/kohya-flux-lora-110293257
Eflux1-dev.safetensors
Fok that is accurate. what is your T5 model name and size?
EI have tried the combined version: t5xxl_fp16.safetensors
and the non combined version from the blacklabs repo:
and the non combined version from the blacklabs repo:
Fplease also apply accelerator setup above and restart and try again
E
Fthis is wrong
Fyou need to use this one
Fcomfyanonymous/flux_text_encoders
Ekk will grab and retest
Fyou know i have model downloader too 
EExtra note I do have a 4090 and a 3080 plugged in to the same rig. Though I usually only use 4090 for training 3080 is mainly there for other uses. So I normally set accelerator to only use GPU ID 0
Fi have rtx 3090 and 3060
Fshouldnt be a problem
Edo you have both enabled in accelerator ? I thought it would cause the 3060 vram to max out and crash the training?
Fnope it uses only single gpu
Fduring training
Funless you set multiple gpus
EKk so what I though so weird
EI am gonna run your model downloader now incase one of my models is the issue
Fclip l arrived
F
EAnswer - VAE was the issue -_-
EI was using the one from BlackLabs but it wasnt encoding the image correctly for caching
SI hope it’s even effective with flux. I’ve read mixed things about it, curious to your findings
Fusing downloader is best :d
Fi plan to train today
Nwith latest build, didn't work eather
Fwhich config you tried
Nepochs 50, repeats 1, size 512,512. Tried v5 slow
Fit is impossible
Fit should use lesser than 10 gb
Fwhat is your batch size
N
Fsend me your json file as prrivatre message
Fyou are not using lora tab
Fnow extract config from zip file again
Ffresh
Fload into lora tab fresh
