also dont upgrade torch 2.5 it may use more vram and it didnt bring speed when using single layer st
also dont upgrade torch 2.5 it may use more vram and it didnt bring speed when using single layer strategy
Fplease do this and see what error you get
F
Fopen google remote
Fi will pin point your error
Fif you have 1 time loaded config into dreambooth it will get corrupted
LI haven't updated anything. I have RAM running at max level, before moving text encoders to gpu and after that it gives an error
Fprivate message me your config json file please
Falso show your task manager gpu before starting training
Fscreenshot of it
M1024 Speeds on windows finally working faster 
Fawesome speed
Fused torch 2.5?
Mjust pulled the latest dev one
F4090 is a beast and shines with flux
Fsdxl had no such speed diff
Fwhen training
Mhave you tried h100?
M5000x series should be a good upgrade for you
Mbut sadly only 28gb vram
Mshould of at least given 32gb vram
Mone trainer results are quite amazing so far, sadly the loras dont work in forge or comfy yet
M@Dr. Furkan Gözükara also all your configs have a bug in them
Fnah i am gonna buy what people cna buy :d
Fwhich one onetrainer?
Mit finally got fixed today, it was actually passing through conv dim and alpha
Mfor flux
M" from bmalsais" this is what he told me
"Well, the GUI was processing some of the values... but it should not have been. So I just took them out of the flus network list of possible parameters. This ensure only the ones officially supported by sd-scripts are used."
"Well, the GUI was processing some of the values... but it should not have been. So I just took them out of the flus network list of possible parameters. This ensure only the ones officially supported by sd-scripts are used."
Mplease check your jsons, they all had conv dim and alpha in them
"constrain": 0,
"conv_alpha": 1,
"conv_block_alphas": "",
"conv_block_dims": "",
"conv_dim": 1,
"constrain": 0,
"conv_alpha": 1,
"conv_block_alphas": "",
"conv_block_dims": "",
"conv_dim": 1,
Meven though it was training from the toml it creates, there was a bug pushing conv alphas and dims, and other unsupoorted params to flux
Fit uses toml there is no other way to push them
Fit was using default of kohya ss when not pushed
Mi confirmed it with kohya and bmaltis, it still passes through json params
Mwell its fixed now
Fi am gonna train again now with latest version
Flets see if any difference there will be
Mwait
Mdont load in your config
Mcreate a new one manually
Fok
Mlet me know if you get better results, i have been getting better, after running loads of tests
Fi will compare with grid lets see :d
Fanything else to test?
Fi got 8 gpus waiting :d
Fi tested text encoder and i got down to 2 text encoder params
Mi am able to stylize my trainings now
Mhave you tested prior loss weight for reg images? also just training it as person instead of man or woman?
Mhave you also tried kohyas recomendation of shift and 3.1582?
Fyep i tested all prior loss
Fgood one gonna test
