but it's better to convert it to fp8? 🙂
but it's better to convert it to fp8? 
Fnot sure. i keep loras as fp32 and base models fp16 
Fswarmui is able to auto cast into used precision
�you must have a lot of diskspace
FHi Doctor ! @Furkan Gözükara SECourses Please.. do you have some tutorial about training sd3.5? I prefer it to flux for some kind of generations.. I feel it more creative when subjects are not real humans ....
Fthank you
Fas a separate issue, my swarm generates another folder with the name stableswarm so that the model installer for windows does not work well because it generates another parallel swarm folder with another new folder of models ... I think that maybe something is not quite right or does not fit with my folder configuration.
Fi plan to do but i couldnt find time for it yet
Fyep
Fi am not sure but developer replying
F
Zhey quick question, im trying to train a LoRa on Flux, I have a 3060 with 12gb so im using your config tier 2 for and following the youtube tutorial. I disabled all my start ups and my task manager is showing im using about 3-400mb of memory on my gpu when I start training. Ive reduced my resolution to 512x512, increased block swaps, and tried using the 10gb config. everytime I get a CUDA out of memory error and the training fails
Zany advice?
Adoes anybody knows if the free kaggle notebook is still good (I'd rather say:"working") at SDXL training? And, if so, a tutorial to use it with the new Kohya interface?
TGuys, are the current commit of sd3-flux.1 branch kohya_ss working for you? I swear the branch worked perfectly yesterday, but I'm having this annoying error.
Tdid 2 clean installs with pip cache purging to no avail
Talso this warning, not shure if they are new or normal
TNever mind, I'm just an idiot and forgot --noverify flag, so torch and xformers version reset on gui.sh run. Here goes an hour of my life 
D
Tthis is different error, it says CUDA is out of memory, so it probably failed to fit in VRAM whatever you are trying to train
Tyou can use nvitop tool to monitor your GPU status, ensure you have enough of free VRAM to fit all models (flux + t5xxl + clip_l + vae)
Taround 29 GB for fp16 flux dev
DI used the config that uses 8400mb of vram. I have 12gb 4070 super that had around 10gb free vram. I checked that using nvitop
DAlso, I'm thinking about doing it on massed compute instead. In the tutorial, he said that kohya is not on main branch there. Can I use/switch to this sd3-flux.1 on massed compute?
TI would say doing training in the cloud is the best option unless you can't afford spending a few bucks on this project. It requires more effort though and basic linux administration skills
Fit works but i cant say sdxl very good
Fhow much RAM you have?
Fyou should upgrade 64 GB and it will fix your issue 100%
Qcan i run multiple instances of Khoya - on different GPUs for dreambooth finetuning?
Qif i install 6 instances on C:\Ko1
C:\Ko2 etc
C:\Ko2 etc
Qi set them each to use a different GPU, should that work?
How can i set the port for Kohya?
How can i set the port for Kohya?
Fyou dont need that
Fyou can set gpu id in accelerate tab
Fjust start the kohya multiple times
Qstart the same Kohya multiple times
Qand set each on different GPU
Fyes
Fmake sure to use different image folders
Fsince each cache overwrites others
QHow can i do multiple people on a single finetune
Fnot possible yet it bleeds
Fbut if you have them in same picture
Fthen model can learn
Qhow can i then mark them with the right token
QHappy to run a few experiments to see if what works what doesnt
Flike
Fohwx woman bbuk man and some random rare word for third and so on
