Software Engineering Courses (SECourses)•17mo ago

1024 Speeds on windows finally working faster 🙂

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?

MaxiviOP•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

MaxiviOP•9/1/24, 12:01 AM

have you tried h100?

MaxiviOP•9/1/24, 12:02 AM

5000x series should be a good upgrade for you

MaxiviOP•9/1/24, 12:02 AM

but sadly only 28gb vram

MaxiviOP•9/1/24, 12:02 AM

should of at least given 32gb vram

FFurkan Gözükara SECourses 4090 is a beast and shines with flux

MaxiviOP•9/1/24, 12:03 AM

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

MaxiviOP•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?

MaxiviOP•9/1/24, 12:05 AM

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

MaxiviOP•9/1/24, 12:05 AM

for flux

MaxiviOP•9/1/24, 12:05 AM

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

GitHub

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

MaxiviOP•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."

MaxiviOP•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,

MaxiviOP•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

MaxiviOP•9/1/24, 12:09 AM

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

MaxiviOP•9/1/24, 12:09 AM

well its fixed now

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