are there strength and weaknesses for them or do most just use BLIP?
are there strength and weaknesses for them or do most just use BLIP?
Fyes each one fits
Fdifferent cases
Fi try them
JHello Dr. Furkan, i followed your tutorial to get a lora trained using a rtx 4090 gpu, but at my studio we are wanting to use multi-gpu training for sdxl, ive run into issues trying to configure kohya ss to do this properly (it ends up running into a memory issue after maxing out the memory of the first gpu), would you be willing to help me with this issue?
Fi am giving private consultation and i can help
Fbut gpus are on same machine right?
Jyea theyre on the same machine
Fthen i can help
JThank you! when can we talk about it?
Fi am available for several hours more from now on
JJ@Furkan Gözükara SECourses These are the configure settings I use, am I doing anything wrong here to where I will hit a memory road block while training?
Flooking good
J@Furkan Gözükara SECourses here are my training settings
Fsadly i dont have best settings for SD 1.5
Ji was using your settings for sdxl lora training
Fah yes they are decent but not very best
Fbest is SDXL DreamBooth
Joh i see
JFyour install selections
For better
Fdelete your accelerator settings
Fit shows at the end where it is saved
Fdelete this
Jill try that, thank you
J@Furkan Gözükara SECourses hmm i still seem to be getting the error
Fyou have 3 gpus?
Jyes
Fwell your setup wrong
Fi can professionally help
For you can limit to 1 gpu
Feasily
Fset CUDA_VISIBLE_DEVICES=0
Fadd this to .bat file
Jyea, i understand how to limit it to one gpu, but i would like to set it up for 3 gpu's
Fare your gpus same?
Jis that do-able?
Fwhat are they
Jtheyre all 4090's
Fye i think should work
Jwhat would i need to change to get that to work?
FDon't know. I have to connect your pc and figure out
Jim not sure that my workplace would be comfortable with that, we deal with a lot of proprietary information, is there any other way to go about it?
L@Josh It's the pytorch package I think, all the libraries in python are a hellish mess.
You can try to find a package that contains the distributed compilation code, you can set your configuration to single_gpu or you can move it into WSL.
I think going WSL is the best option
You can try to find a package that contains the distributed compilation code, you can set your configuration to single_gpu or you can move it into WSL.
I think going WSL is the best option
JLYou use Win11 ?
Lyou can google for pytorch, windows, nccl .. maybe you find a working package or you can compile it.
Or on Windows 11 you install WSL-2 with Ubuntu or Debian and then use some linux/WSL tutorial. It should be easier to train that way and faster - nvidia cuda drivers work better in the linux environment
Or on Windows 11 you install WSL-2 with Ubuntu or Debian and then use some linux/WSL tutorial. It should be easier to train that way and faster - nvidia cuda drivers work better in the linux environment
Fi dont think so necessary
