cant linux has shared vram?
cant linux has shared vram?
MOk I could maybe create a ticket on it's git
MI search but didn't find a way to do that it seem not implemented in linux nvidia driver
Fplease make a detailed reply here and tag @kohya-ss : https://github.com/kohya-ss/sd-scripts/pull/1374
GitHub
Replace SD3Tokenizer with the original CLIP-L/G/T5 tokenizers.
Extend the max token length to 256 for T5XXL.
Refactor caching for latents.
Refactor caching for Text Encoder outputs
Extract arch...
Extend the max token length to 256 for T5XXL.
Refactor caching for latents.
Refactor caching for Text Encoder outputs
Extract arch...
Fthere are also other expers some of the may chime in
Mok thank you I will do that
Vdoes it take a few mins in smarmui on kaggle for flux model?
Vto create 1 image, I mean? seeming very slow
Fye it is slow. kaggle doesnt give good gpus :/
Fi wish they started giving bf16
Fwe could trian flux there 
Fi trained 24 images on 4x gpu and even 240 epoch is not overtrained. this is impact of effective batch size increase
JHi there, may I ask you something, for kohya_ss, for generate the meta_lat.json for the full finetuning of flux, do you know which path of the model should I set as argument, the path to flux1-dev.safetensors seems to not working
Fi didnt research it yet. hopefully next week
Vyes I am able to do use swarmUI now, thanks a lot
JWhats the best tutorial for Flux image to image generation for product images.?
Fi never done that but i have shown image to image
Fbut not for flux
D@Dr. Furkan Gözükara new test with high TE learning rate, now with unique tokens no class and person class regularization, I see that using person class avoid class mixing
Dthese are my regularizations
Dworks much better if I use man class it overtrains the class man, en mix them all
Dis a couple of thousands of regularization pictures
Fdo you see improved results?
Fi am about to compare 800 images generation
Fwaiting grid to finish
F780 - each one 48 steps FLUX image
Dyes but is a preliminary test, I stopped and restarted the training with different settings a couple of times
Dnow I started the real test
DI will report in a few hours
Dremember my goal is to train multiple people no concept or class bleading
Dlest see if I succeed
Fgreat
DIn my last test I wasn't using unique tokens and I changed tokens and removed class in the middle of a training resuming..
DIt was a mess
Dother note, all my regularization are captioned
D27hs left
Fnice. is this single gpu?
Drtx 3090
Dmultiple gpu can be used on kohya?
Fok it is really good speed
Fof course
Fi recorded tutorial
Fyou can even use 8 gpus :d
Fediting the tutorial
DI have just one
DFFyou can rent on massed compute
F1.25$ per hour 4x A6000 GPU
Fok i am gonna update configs with clip l
Fthere is slight improvement
