Software Engineering Courses (SECourses)•11mo ago

Hi @Furkan Gözükara SECourses, I've tried both `flux1-dev-fp8.safetensors` and `flux1-dev.safetensor

Hi @Furkan Gözükara SECourses,
I've tried both

flux1-dev-fp8.safetensors

flux1-dev-fp8.safetensors

flux1-dev-fp8.safetensors

flux1-dev-fp8.safetensors and

flux1-dev.safetensors

flux1-dev.safetensors

flux1-dev.safetensors

flux1-dev.safetensors based on

8GB_GPU_Quality_Tier_4_FP8_6850MB_13.2_Second_IT.json

8GB_GPU_Quality_Tier_4_FP8_6850MB_13.2_Second_IT.json

8GB_GPU_Quality_Tier_4_FP8_6850MB_13.2_Second_IT.json

8GB_GPU_Quality_Tier_4_FP8_6850MB_13.2_Second_IT.json config
on 48 images with faces pre-captioned with joy captioning and replaced with

ohwx

ohwx

ohwx

ohwx token,
it takes too much time - estimated 45 days! is that make sense??
Shared GPU raised to 14.2 and GPU RAM is on 7.5 GB.
I've attached all screenshots

GGuy Bertental Hi @Furkan Gözükara SECourses, I've tried both `flux1-dev-fp8.safetensors` and `...

Guy BertentalOP•2/21/25, 4:42 PM

another screenshot with more lines from training console

NNeuromasters You want to share? I'll be sure to test it And here's a question, has `Hunyuan3D...

Furkan Gözükara SECourses•2/21/25, 5:48 PM

i will share hopefully after finalizing

V3ldrin•2/21/25, 5:49 PM

Hi, I also have a problem when training with kohya: I have a geforce rtx 3900 with 24 vram, I use the models from your youtube tutorial video (flux dev, fp16, etc.etc). Before training, I clear vram (restart, autostart apps off, etc). When I check, I have enough vram under 500 mb. When I start training, it goes up to 23 GB and ALSO the shared Vram goes up. Why? Is that normal? using the 24GB_GPU_Quality_Tier_3_23150MB_5.9_Second_IT

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 Kohya implemented this - https://github.com/kohya-ss/sd-scripts/pull/1864 can yo...

Furkan Gözükara SECourses•2/21/25, 5:51 PM

i dont plan to research this atm

Jjulius trying to generate 4 images at the same time (using 4 L40) but it seems like it ...

Furkan Gözükara SECourses•2/21/25, 5:51 PM

click load model

Furkan Gözükara SECourses•2/21/25, 5:51 PM

it will then use all 4

Furkan Gözükara SECourses•2/21/25, 5:51 PM

or generate like 16

GGuy Bertental Hi @Furkan Gözükara SECourses, I've tried both `flux1-dev-fp8.safetensors` and `...

Furkan Gözükara SECourses•2/21/25, 5:52 PM

your folder repeating is 40

Furkan Gözükara SECourses•2/21/25, 5:52 PM

therefore you are doing 392k steps

Furkan Gözükara SECourses•2/21/25, 5:52 PM

make folder name like 1_ohwx man

Furkan Gözükara SECourses•2/21/25, 5:52 PM

your speed is good

Furkan Gözükara SECourses•2/21/25, 5:52 PM

10 second / it

VV3ldrin Hi, I also have a problem when training with kohya: I have a geforce rtx 3900 wi...

Furkan Gözükara SECourses•2/21/25, 5:52 PM

totally normal

Furkan Gözükara SECourses•2/21/25, 5:52 PM

you can increase block swap count

Furkan Gözükara SECourses•2/21/25, 5:53 PM

and compare speed

FFurkan Gözükara SECourses i dont plan to research this atm

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•2/21/25, 5:53 PM

Okay but can you tell me how I can run this in kohya? I am not able to understand.

V3ldrin•2/21/25, 5:53 PM

okay. thanks. it trains like 6-7 hours

V3ldrin•2/21/25, 5:53 PM

and works

V3ldrin•2/21/25, 5:53 PM

was just wondering

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 Okay but can you tell me how I can run this in kohya? I am not able to understan...

Furkan Gözükara SECourses•2/21/25, 5:53 PM

i dont know either didnt look at it

VV3ldrin okay. thanks. it trains like 6-7 hours

Furkan Gözükara SECourses•2/21/25, 5:53 PM

totally good

Furkan Gözükara SECourses•2/21/25, 5:54 PM

unless you have rtx 5090 you have to use shared vram

FFurkan Gözükara SECourses i dont know either didnt look at it 🙂

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•2/21/25, 5:54 PM

Please take a look whenever you get time.

AAiInfluence it's the best currently what do you mean they actually did remove ??

Timson•2/21/25, 6:02 PM

daily credits removed from all plans, only monthly remained, effective for new subs after 20 feb

TTimson daily credits removed from all plans, only monthly remained, effective for new s...

AiInfluence•2/21/25, 6:02 PM

this is bad wow

AiInfluence•2/21/25, 6:03 PM

But as I said I think they abuse monopoly rn

Timson•2/21/25, 6:04 PM

I don't think even current monetisation is profitable for them. video generation compute is crazy expensive

AiInfluence•2/21/25, 6:04 PM

is it ? For 5seconds and from ab image frame?

Timson•2/21/25, 6:04 PM

yeah, 5 seconds at 30 fps is 150 frames each taking unknown number of steps for diffusion model

Timson•2/21/25, 6:05 PM

we are talking many minutes of a100 probably per generation

AiInfluence•2/21/25, 6:05 PM

so 0.50$

Timson•2/21/25, 6:07 PM

So you get 660 monthly credits for 80$ yearly, 35 per 5 s generation, so you pay ~0.35$ per gen for their top model 1.6 in pro mode

Timson•2/21/25, 6:07 PM

without daily credits

Timson•2/21/25, 6:08 PM

a ton of freeloaders with free plan

AiInfluence•2/21/25, 6:12 PM

I give it 2 months for a better open source model to come

AiInfluence•2/21/25, 6:12 PM

Omni human and Goku are on their way too so

Darth Teethius•2/21/25, 10:55 PM

for the sort by similarity app, what kind of speeds should i be expecting? right now it takes about an hour to process 20 images

DDarth Teethius for the sort by similarity app, what kind of speeds should i be expecting? right...

Furkan Gözükara SECourses•2/21/25, 11:42 PM

it is wrong

Furkan Gözükara SECourses•2/21/25, 11:42 PM

it means not using your gpu

Furkan Gözükara SECourses•2/21/25, 11:42 PM

it uses tensorflow so it is extremely delicate to the cuda and cudnn libraries

Furkan Gözükara SECourses•2/21/25, 11:42 PM

in your system

Furkan Gözükara SECourses•2/21/25, 11:42 PM

it sucks from this perspective actually

Darth Teethius•2/21/25, 11:56 PM

yeah its all on the cpu atm. not sure why

Darth Teethius•2/21/25, 11:57 PM

i have all the prereqs installed. i think it might be defualting to a cached version of keras and protobuf that aren't compatible

Darth Teethius•2/21/25, 11:57 PM

thank you. might be time for a clean install again lol

DDarth Teethius thank you. might be time for a clean install again lol

Furkan Gözükara SECourses•2/22/25, 12:10 AM

it is related to cuda and cudnn 99%

Furkan Gözükara SECourses•2/22/25, 12:17 AM

i plan to move cuda 12.8 but torchvision and torchaudio cu128 still not ready. after that hopefully every app can work on 5000 series as well

Furkan Gözükara SECourses•2/22/25, 11:52 AM

wow this week was so busy still reading all emails

Hi @Furkan Gözükara SECourses, I've tried both `flux1-dev-fp8.safetensors` and `flux1-dev.safetensor

Similar Threads