Software Engineering Courses (SECourses)•17mo ago

Hmm, I can see the 12gb it looks like it is failing pretty early. Do you have a game running in the

Hmm, I can see the 12gb it looks like it is failing pretty early. Do you have a game running in the BG or another stable diffusion UI running in the BG.

It looks like it is failing because there is only 5gb free on the card and it needs 6gb

EEvivrus Hmm, I can see the 12gb it looks like it is failing pretty early. Do you have a ...

LW24Ai•8/31/24, 11:06 PM

No, I don't have any other background tasks, I only work with Kohya.

LW24Ai•8/31/24, 11:10 PM

LW24Ai•8/31/24, 11:23 PM

I don't know, I've been testing everything and something's going wrong.

SSTUFCS SUPER !

Furkan Gözükara SECourses•8/31/24, 11:40 PM

sorry for delay. i was planning actually today too but i got some real life bussines. moving from moutain house to city again

GGomezBro Link to site, coupon and the multi GPU video when you drop it. I will do this im...

Furkan Gözükara SECourses•8/31/24, 11:40 PM

sure showna ll

GGomezBro What do you estimate time/cost would be for 1000 images?

Furkan Gözükara SECourses•8/31/24, 11:40 PM

depends on how many epochs

Furkan Gözükara SECourses•8/31/24, 11:40 PM

tell me i can tell you

Mmacba @Dr. Furkan Gözükara and colleagues, have you seen this? https://github.com/bg...

Furkan Gözükara SECourses•8/31/24, 11:40 PM

ye i am aware but i still like kohya

LLW24Ai You tricked me. The 23gb FLUX dev FP16 with preset Rank_5_11498MB_Slow does not ...

Furkan Gözükara SECourses•8/31/24, 11:42 PM

i train in my rtx 3060

Furkan Gözükara SECourses•8/31/24, 11:43 PM

how much vram your system using before starting the script?

Furkan Gözükara SECourses•8/31/24, 11:43 PM

also dont upgrade torch 2.5 it may use more vram and it didnt bring speed when using single layer strategy

AAIGambino Dr When I click windows install step 1 it just opens and closes

Furkan Gözükara SECourses•8/31/24, 11:43 PM

please do this and see what error you get

Furkan Gözükara SECourses•8/31/24, 11:43 PM

LLW24Ai You tricked me. The 23gb FLUX dev FP16 with preset Rank_5_11498MB_Slow does not ...

Furkan Gözükara SECourses•8/31/24, 11:44 PM

open google remote

Furkan Gözükara SECourses•8/31/24, 11:44 PM

i will pin point your error

LLW24Ai I don't know, I've been testing everything and something's going wrong.

Furkan Gözükara SECourses•8/31/24, 11:44 PM

if you have 1 time loaded config into dreambooth it will get corrupted

FFurkan Gözükara SECourses also dont upgrade torch 2.5 it may use more vram and it didnt bring speed when u...

LW24Ai•8/31/24, 11:46 PM

I haven't updated anything. I have RAM running at max level, before moving text encoders to gpu and after that it gives an error

LLW24Ai I haven't updated anything. I have RAM running at max level, before moving text ...

Furkan Gözükara SECourses•8/31/24, 11:47 PM

private message me your config json file please

Furkan Gözükara SECourses•8/31/24, 11:47 PM

also show your task manager gpu before starting training

Furkan Gözükara SECourses•8/31/24, 11:47 PM

screenshot of it

Maxivi•9/1/24, 12:00 AM

1024 Speeds on windows finally working faster

MMaxivi 1024 Speeds on windows finally working faster 🙂

Furkan Gözükara SECourses•9/1/24, 12:01 AM

awesome speed

Furkan Gözükara SECourses•9/1/24, 12:01 AM

used torch 2.5?

Maxivi•9/1/24, 12:01 AM

just pulled the latest dev one

Furkan Gözükara SECourses•9/1/24, 12:01 AM

4090 is a beast and shines with flux

Furkan Gözükara SECourses•9/1/24, 12:01 AM

sdxl had no such speed diff

Furkan Gözükara SECourses•9/1/24, 12:01 AM

when training

Maxivi•9/1/24, 12:01 AM

have you tried h100?

Maxivi•9/1/24, 12:02 AM

5000x series should be a good upgrade for you

Maxivi•9/1/24, 12:02 AM

but sadly only 28gb vram

Maxivi•9/1/24, 12:02 AM

should of at least given 32gb vram

FFurkan Gözükara SECourses 4090 is a beast and shines with flux

Maxivi•9/1/24, 12:03 AM

one trainer results are quite amazing so far, sadly the loras dont work in forge or comfy yet

Maxivi•9/1/24, 12:04 AM

@Dr. Furkan Gözükara also all your configs have a bug in them

MMaxivi have you tried h100?

Furkan Gözükara SECourses•9/1/24, 12:04 AM

nah i am gonna buy what people cna buy :d

MMaxivi @Dr. Furkan Gözükara also all your configs have a bug in them

Furkan Gözükara SECourses•9/1/24, 12:05 AM

which one onetrainer?

Maxivi•9/1/24, 12:05 AM

it finally got fixed today, it was actually passing through conv dim and alpha

Maxivi•9/1/24, 12:05 AM

for flux

Maxivi•9/1/24, 12:05 AM

https://github.com/bmaltais/kohya_ss/commit/8a524c7bae9e070ad69deb0389b3b8bc46fc495e

GitHub

Removed unsupported parameters from flux lora network · bmaltais/ko...

Maxivi•9/1/24, 12:05 AM

" 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."

Maxivi•9/1/24, 12:07 AM

please 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,

Maxivi•9/1/24, 12:08 AM

even though it was training from the toml it creates, there was a bug pushing conv alphas and dims, and other unsupoorted params to flux

MMaxivi even though it was training from the toml it creates, there was a bug pushing co...

Furkan Gözükara SECourses•9/1/24, 12:08 AM

it uses toml there is no other way to push them

Furkan Gözükara SECourses•9/1/24, 12:08 AM

it was using default of kohya ss when not pushed

Maxivi•9/1/24, 12:09 AM

i confirmed it with kohya and bmaltis, it still passes through json params

Maxivi•9/1/24, 12:09 AM

well its fixed now

Furkan Gözükara SECourses•9/1/24, 12:09 AM