Software Engineering Courses (SECourses)•17mo ago

Hey! I'm trying torch 2.5 to see if I can speed things up. Think I set it up properly. But I'm not s

Hey! I'm trying torch 2.5 to see if I can speed things up. Think I set it up properly. But I'm not seeing much improvement [some, very little].

25000 steps / 66 hours on a RTX 3090 | 25 images / 5 repeats / 200 epochs

Is that normal for Flux?

Uuisato Hey! I'm trying torch 2.5 to see if I can speed things up. Think I set it up pro...

dsienra•8/30/24, 4:00 AM

how many it/s?

dsienra•8/30/24, 4:08 AM

why are you using 5 repeats? 5 repeats 200 epochs = 1000 epochs is excessive

Jackie•8/30/24, 4:10 AM

The real photos that I used for training

Ddsienra how many it/s?

uisatoOP•8/30/24, 4:14 AM

9.47

Ddsienra why are you using 5 repeats? 5 repeats 200 epochs = 1000 epochs is excessive

uisatoOP•8/30/24, 4:14 AM

how much'd you suggest? for a style LORA, first test

Uuisato 9.47

dsienra•8/30/24, 4:18 AM

slow

dsienra•8/30/24, 4:19 AM

what graphics card?

Ddsienra what graphics card?

uisatoOP•8/30/24, 4:20 AM

3090

dsienra•8/30/24, 4:22 AM

with T5 attention mask off I get 4.5 it/s in my 3090 ti

dsienra•8/30/24, 4:23 AM

with T5 attention mask enabled 6.5 it/s

Uuisato how much'd you suggest? for a style LORA, first test

dsienra•8/30/24, 4:24 AM

I suggest 1 repetition 200 epochs

uisatoOP•8/30/24, 4:24 AM

would you suggest 1 repeat / 200 epochs? 2 repeats / 100 epochs?

uisatoOP•8/30/24, 4:24 AM

great.

uisatoOP•8/30/24, 4:25 AM

thanks!

dsienra•8/30/24, 4:25 AM

5000 steps

Ddsienra with T5 attention mask off I get 4.5 it/s in my 3090 ti

uisatoOP•8/30/24, 4:25 AM

have u seen quality differences with / without?

dsienra•8/30/24, 4:27 AM

no in ma case but is better if you have captions

dsienra•8/30/24, 4:27 AM

theoretically

uisatoOP•8/30/24, 4:27 AM

so if i'm using captions, i should aswell use t5 attention

dsienra•8/30/24, 4:28 AM

yes, you can try with and without

dsienra•8/30/24, 4:30 AM

for a firts run you can try the fast way 1 repeat 150 or 200 epochs and see the results

dsienra•8/30/24, 4:30 AM

then try with attention mask and compare

dsienra•8/30/24, 4:31 AM

just a suggestion, there is no one solution for every case

dsienra•8/30/24, 4:33 AM

at this point is trial an error

AIGambino•8/30/24, 4:38 AM

Any way to improve these settings for ai toolkit?

message.txt2.23KB

AIGambino•8/30/24, 4:39 AM

102 images so far has been going 12 hours

AAIGambino Any way to improve these settings for ai toolkit?

dsienra•8/30/24, 4:58 AM

I only used kohya

dsienra•8/30/24, 4:58 AM

but you are using adamw8bit

dsienra•8/30/24, 4:59 AM

it consumes more vram

dsienra•8/30/24, 4:59 AM

I use adafactor

dsienra•8/30/24, 5:00 AM

what graphic card do you have?

mikemenders•8/30/24, 5:04 AM

Anyone use Prodigy for Flux traning?

Mmikemenders Anyone use Prodigy for Flux traning?

dsienra•8/30/24, 5:05 AM

not me, it too memory hungry

SeatownSin•8/30/24, 5:28 AM

are there any differences i should be aware of when making a gguf of a flux.1-dev fp32 safetensor model, compared to the way i would do it for an llm? just looking to make a q8 for people with less vram.

Tacite•8/30/24, 6:14 AM

https://huggingface.co/mo137/FLUX.1-dev_Q8-fp16-fp32-mix_8-to-32-bpw_gguf/tree/main

mo137/FLUX.1-dev_Q8-fp16-fp32-mix_8-to-32-bpw_gguf at main

TTacite https://huggingface.co/mo137/FLUX.1-dev_Q8-fp16-fp32-mix_8-to-32-bpw_gguf/tree/m...

J2_Nedo•8/30/24, 6:38 AM

using only original ones from Black forrest labs (Flux1-dev),
there are too many versions out for me to keep track of and whether they are all
compatible with my LoRAs or not.

Tacite•8/30/24, 6:57 AM

Yeah they are original, you can read the repo page to have the info