Software Engineering Courses (SECourses)•16mo ago

nevermind it just started training although extremely slowly. If the lora source images, and target

nevermind it just started training although extremely slowly. If the lora source images, and target direectory are on a slower drive would that cause a delay?

WWebsteria nevermind it just started training although extremely slowly. If the lora source...

Furkan Gözükara SECourses•9/5/24, 3:08 PM

it can cause delay in initial

Furkan Gözükara SECourses•9/5/24, 3:08 PM

start

Furkan Gözükara SECourses•9/5/24, 3:08 PM

but not after it

lilwashuu10•9/5/24, 3:08 PM

Im curious how much RAM are you running with, looks like FLUX is heavy on the PC RAM as well

Llilwashuu10 Im curious how much RAM are you running with, looks like FLUX is heavy on the PC...

Furkan Gözükara SECourses•9/5/24, 3:12 PM

ye it uses a lot

Furkan Gözükara SECourses•9/5/24, 3:12 PM

i have 64 gb so never cared :d

Furkan Gözükara SECourses•9/5/24, 3:12 PM

also you can set virtual ram

Furkan Gözükara SECourses•9/5/24, 3:12 PM

ram is nowadays easier

lilwashuu10•9/5/24, 3:13 PM

Yeah I have more coming in the mail, but I think paging is what is causing some peoples slow down

LuckyStrike•9/5/24, 3:14 PM

Hey, so I ran a couple of successful Flux lora's. with the rank three settings, but. the file size is like 2.3 gigs and I'm thinking I must have done something wrong to end up with that file size. I remember having that mistake before when I accidentally ran it on the Dreambooth tab instead of the Lora tab. But that's not what I did here. Is there anything apparent that I'm doing wrong?

lilwashuu10•9/5/24, 3:14 PM

From the video it looked like Furkans where about that size

LuckyStrike•9/5/24, 3:15 PM

Oh, so maybe that's the expected size using this approach. It's big, but if it works, then I'm totally fine with that.

lilwashuu10•9/5/24, 3:16 PM

Yeah I tried the rank 12 which does i think 512x512 training and the model was 525MB

lilwashuu10•9/5/24, 3:17 PM

I am sure it goes up in size with the resolution, since it is gathering more information

FFurkan Gözükara SECourses rank 3 is for 24 gb

magical•9/5/24, 3:18 PM

is rank 3 config bf16?

Mmagical is rank 3 config bf16?

Furkan Gözükara SECourses•9/5/24, 3:32 PM

all are bf16 but rank 5 and lower cast into fp8

Furkan Gözükara SECourses•9/5/24, 3:37 PM

some 16 gb configs doesnt cast but they are lower resolution

magical•9/5/24, 3:39 PM

I was using this extension https://github.com/cheald/sd-webui-loractl with a1111 to change the loras weight at different steps for example I can start one lora at first step and stop it at 15th step and start another lora at 15th step until the end. Now I'm looking if I can do it in SwarmUI and the only thing I found is Lora Block Weight node in ComfyUI-Inspire-Pack.

magical•9/5/24, 3:39 PM

So could I download this https://github.com/ltdrdata/ComfyUI-Inspire-Pack and use it inside SwarmUI?

Mmagical So could I download this https://github.com/ltdrdata/ComfyUI-Inspire-Pack and us...

Furkan Gözükara SECourses•9/5/24, 3:58 PM

not sure may not work

slipkn67•9/5/24, 4:21 PM

it has been running over 3hrs and only 844/2800 done. How do I pause it to continue later? Running 14 photos on RTX 3060 12GB.

JohnDopamine•9/5/24, 4:29 PM

@Dr. Furkan Gözükara You testing T5 training yet?

JJohnDopamine @Dr. Furkan Gözükara You testing T5 training yet?

Furkan Gözükara SECourses•9/5/24, 4:30 PM

didnt have chance yet

Furkan Gözükara SECourses•9/5/24, 4:30 PM

i have been trying to fix broken scripts reply people :d

Sslipkn67 it has been running over 3hrs and only 844/2800 done. How do I pause it to conti...

Furkan Gözükara SECourses•9/5/24, 4:30 PM

sadly you cant

Furkan Gözükara SECourses•9/5/24, 4:30 PM

there is use last checkpoint for kohya but people say not working i didnt test

Furkan Gözükara SECourses•9/5/24, 4:30 PM

look at the tips

Furkan Gözükara SECourses•9/5/24, 4:30 PM

#tips-info

FFurkan Gözükara SECourses sadly you cant

slipkn67•9/5/24, 4:32 PM

ok, so the average speed is OK for my setup? ~4hrs and only 875/2800 on 12GB RTX3060

Sslipkn67 ok, so the average speed is OK for my setup? ~4hrs and only 875/2800 on 12GB RTX...

Furkan Gözükara SECourses•9/5/24, 4:35 PM

20 second / it is best

Furkan Gözükara SECourses•9/5/24, 4:35 PM

you can get with 1024

Furkan Gözükara SECourses•9/5/24, 4:36 PM

tell me your step speed

FFurkan Gözükara SECourses tell me your step speed

slipkn67•9/5/24, 4:36 PM

21 sec/it, avg loss 0.4

Furkan Gözükara SECourses•9/5/24, 4:43 PM

excellent news we got for 16gb gpus

Furkan Gözükara SECourses•9/5/24, 4:43 PM

Sslipkn67 21 sec/it, avg loss 0.4

Furkan Gözükara SECourses•9/5/24, 4:43 PM

it is good

FFurkan Gözükara SECourses Click to see attachment

druhl•9/5/24, 5:29 PM

I believe there were already 2 configs/ ranks (one slow & the other fast) for 16GB GPUs?

slipkn67•9/5/24, 7:10 PM

where do I type my prompt to generate images? Or is Kohya_SS GUI doesn't support it?

magical•9/5/24, 7:12 PM

the learning rate in the flux config with single gpu is 0.00005. with sdxl it was 0.0004. is there a reason why it is much smaller? and does it mean we have to train it twice as much steps? so for example if i was training 4500 steps with sdxl do i have to train more than that because the learning rate is much smaller?

xlads•9/5/24, 7:13 PM

"ERROR: Could not find a version that satisfies the requirement xformers==0.0.27.post2 (from versions: none)
ERROR: No matching distribution found for xformers==0.0.27.post2" Getting this error in very first step for "FLUX LoRA Training Simplified" I ran "pip install xformers" and it says that the xformers installed succsefully but when i try to install the first step again it says the same error

Tacite•9/5/24, 7:15 PM

Can you do pip freeze in your venv?

Tacite•9/5/24, 7:15 PM

And post the result

xlads•9/5/24, 7:20 PM

Im going to uninstall kohya_ss from my pc then reinstall now that i have necessary requirements hopefully it works

xlads•9/5/24, 7:26 PM

i uninstalled "kohya_ss from my pc" after getting this "ERROR: Could not find a version that satisfies the requirement xformers==0.0.27.post2 (from versions: none)
ERROR: No matching distribution found for xformers==0.0.27.post2

[notice] A new release of pip is available: 23.0.1 -> 24.2
[notice] To update, run: python.exe -m pip install --upgrade pip" installing xformers then reintalled it but still getting same error

Xxlads i uninstalled "kohya_ss from my pc" after getting this "ERROR: Could not find a ...

slipkn67•9/5/24, 7:30 PM

bro I had exact same problem and Gozukara replied it's fine

slipkn67•9/5/24, 7:30 PM

qwen-flux-wan-such

magical•9/5/24, 7:32 PM

can we use trained flux loras in kaggle?

xlads•9/5/24, 8:02 PM

"To create a public link, set

share=True

share=True

launch()

launch()

.
ERROR: Exception in ASGI application
Traceback (most recent call last):
File "C:\Kohya_GUI_Flux_Installer_26 (1)\kohya_ss\venv\lib\site-packages\pydantic\type_adapter.py", line 279, in _init_core_attrs
self._core_schema = _getattr_no_parents(self._type, 'pydantic_core_schema')
File "C:\Kohya_GUI_Flux_Installer_26 (1)\kohya_ss\venv\lib\site-packages\pydantic\type_adapter.py", line 121, in _getattr_no_parents
raise AttributeError(attribute)
AttributeError: pydantic_core_schema

During handling of the above exception, another exception occurred:" getting this error when i click on the file inside of Kohya_ss configuration tab folder and the ui is also showing me "error
Connection errored out."

xlads•9/5/24, 8:08 PM

Anyone know what it could be?

Ddruhl I believe there were already 2 configs/ ranks (one slow & the other fast) for 16...

Furkan Gözükara SECourses•9/5/24, 8:29 PM

yes but they had lower quality

nevermind it just started training although extremely slowly. If the lora source images, and target

Similar Threads