Alright will try, it's just that this isn't mentioned anywhere, or did I miss it??
Alright will try, it's just that this isn't mentioned anywhere, or did I miss it??
Fit is for saving space not mandatory
Ffloat is best quality
ZThere is no option to select FP8... only FP16, BF16, Float, and None.
ZThe 11GB size of LoRA… that’s why it’s better in quality. There’s no magic here. If LoRA weighed 1GB while maintaining quality, that would be a real breakthrough. But models with a size of 23GB, where LoRA doesn’t work properly, and extracted LoRAs at 11GB, are a significant negative effect that needs to be known in advance.
Fok then you need 2 steps
ZWhich steps?
Fconvert to fp8
FPatreon
Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon
ZIt reduced the size, but the image is completely black.
Z
Z@Furkan Gözükara SECourses This doesn’t work. I performed LoRA extraction from the model with conversion parameters set to FP16 quality, then converted this LoRA to FP8, and again got a black screen.
ZAre you sure your method works?
Fyep working
Flooks like you did accurate maybe something went wrong or swarmui broken etc
NI know there are much better ways to train a face, but I've got some credits I need to burn through on dreamlook.ai - can anyone suggest best settings to use with their expert mode?
Qfor a multi-gpu system Kohya cannot train. It gives me an error for gpu id not set. I tried to specify the GPU id in the Windows_start_Kohya_SS.bat file by somthing like this
set CUDA_VISIBLE_DEVICES=5
changing the device IDs but it didnt start -
running from the Install.bat menu worked for me but it is using the first GPU - it would be great to run paralell if they cannot run across the GPUs
set CUDA_VISIBLE_DEVICES=5
changing the device IDs but it didnt start -
running from the Install.bat menu worked for me but it is using the first GPU - it would be great to run paralell if they cannot run across the GPUs
�@Furkan Gözükara SECourses can a rank 256 fp16 lora work well if I'm using the fp8 version of my flux dev fine tune? or does the lora have to be fp8 as well?
Fas far as i know multi gpu jnot working on windows
Fbut if youy want to train on different gpuıs
Fuse accelerate gpu id setting
Fit should work
TBlack screen for me also after conversion to lora fp8
�in Acceleration you can set GPU, and multi GPU, but for fine tune using multi GPU you need 80 GB GPUs, if I remember correctly. No need to edit the bat file
�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
