im trying to installl kohya on wsl and i did succesfully install however i cannot run a flux train .
im trying to installl kohya on wsl and i did succesfully install however i cannot run a flux train ... dr furkan can you help please
Hidk what you did on the fix file for windows to make it work
NI'd love to figure that out as well - speed gains look pretty sweet. my understanding of the fix file though is it just updates torch, torchvision and xformers (and modifies that app so it doesn't try to rollback those libraries). you should still be able to run without doing the fix step, it would just be slower and use more vram.
KI was planning on installing JoyCaption, but I'm looking in the Windows_Install and it doesn do a git clone from anywhere. It only creates a venv where the batch-file is launched. Usually there is a git clone first, then a directory change and then the rest
KIt's probably in the HF_model_downloader.py, but this way of installing is way messier than the previous instakllations you've done @Dr. Furkan Gözükara
Ncould this be because just the models need to be downloaded, then Doc wrote the app.py and it's in the zip, so nothing to clone?
FMy training takes 40 hours, what should I change to optimize it? enable full bf16 training.
enable fp8 training.
running training / 学習開始
num train images * repeats / 学習画像の数×繰り返し回数: 61
num reg images / 正則化画像の数: 0
num batches per epoch / 1epochのバッチ数: 61
num epochs / epoch数: 200
batch size per device / バッチサイズ: 1
gradient accumulation steps / 勾配を合計するステップ数 = 1
total optimization steps / 学習ステップ数: 12200
enable fp8 training.
running training / 学習開始
num train images * repeats / 学習画像の数×繰り返し回数: 61
num reg images / 正則化画像の数: 0
num batches per epoch / 1epochのバッチ数: 61
num epochs / epoch数: 200
batch size per device / バッチサイズ: 1
gradient accumulation steps / 勾配を合計するステップ数 = 1
total optimization steps / 学習ステップ数: 12200
Nfewer training images or fewer epochs...
Fit has single app file
Fand download models from hugging face
Hwhich libraries or dependencies i need to install for it and will wsl actually speed up the process if not will it be preferable if i just install a linux system such as mint?
Fnot much you have 61 images
Fprobably lesser epoch
Flike 100
For rent a more powerful gpu and increase batch size
Fon windows i couldnt find yet
Hi see and do u have a .sh file for linux for flux / will the runpod one work?
KSo I make my own folder and put the files into there and run them
Frunpod installer
KOK.
Fi am gonna add batch feature now
Fcoding it
Fso like batch size 8 on each gpu :d
Falso fixing a bug where target saved images are resized versions
K? Isn't JoyCaption only a caption-app?
KOr sorry... Maybe I hijacked your conversation
FOK, I'll keep 200 epochs and switch to 8 GPUs to get the best rendering quality at the end. Sound's good?
FOups I can't not enough A6000
KDoes anyone of you know of a way to quickly see what models are no longer used... They are taking up a shitload of space
Fonly caption
Ffor 8 gpus
Fthe config would change
Fyou need to change LR
Fno way to know 
Fyou can delete all and apps should redownload usually
FI took this 2x RTX A6000 [ALT Config]
14 vCPU, 48 GB RAM, 512 GB storage and I'll use 48_GB_GPUs_v2.json
14 vCPU, 48 GB RAM, 512 GB storage and I'll use 48_GB_GPUs_v2.json
Fnice should work perfectg
Fso now you need to start multi gpu
Fyou can do this train in fp8
Fincrease batch size for each gpu
Flike 2 3 how many it goes
Fcalculate total batch size
Fdivide 2 and multiply LR
FExcuse me but I feel like I'm reading hieroglyphics lol, I'm a newbie in this field, can you please clarify?
Fcant help further here it is complex
Fbut i can give private lecture if you need
Fmulti gpu is another level
