Software Engineering Courses (SECourses)•9mo ago

Thanks! 😄 i've just upgraded my setups from 4090 to 5090 and noticing that I have problems running

Thanks!

i've just upgraded my setups from 4090 to 5090 and noticing that I have problems running pretty much everything I tried before

thanks again! Now i think im getting it

QQ Thanks! 😄 i've just upgraded my setups from 4090 to 5090 and noticing that I h...

Furkan Gözükara SECourses•4/5/25, 10:45 PM

great

HRK•4/6/25, 5:56 AM

Hey, I wonder, has anyone tried / have article or any sort of content about full checkpoint finetuning on Wan ? I really want to get my hands into it but it'd be great to have some references to start working on it

HRK•4/6/25, 5:56 AM

as far as I know only DiffSynth offers the possibility to full checkpoint training on Wan, right ?

HHRK Hey, I wonder, has anyone tried / have article or any sort of content about full...

Furkan Gözükara SECourses•4/6/25, 12:41 PM

good point. i havent checked that yet though. currently musubi is good for lora

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•4/6/25, 2:19 PM

What are the optimal sdxl fine tune and lora configs? Having a hard time finding them. Using massedcompute. Onetrainer or dreambooth? Was thinking of using big love xl 2.5 as base checkpoint. Like the realism and quality of pics there.

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•4/6/25, 2:19 PM

@Furkan Gözükara SECourses

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭What are the optimal sdxl fine tune and lora configs? Having a hard time finding...

Furkan Gözükara SECourses•4/6/25, 3:34 PM

i have dreambooth config

Furkan Gözükara SECourses•4/6/25, 3:34 PM

but not for lora. for lora i recommend do dreambooth and extract

�

🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭•4/6/25, 6:16 PM

https://www.patreon.com/posts/very-best-for-of-89213064 this one?

Patreon

The Very Best Workflow For SDXL DreamBooth / Full Fine Tuning - Res...

Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon

�🍭🎀 𝒜𝓋𝒶 𝐹𝓇𝒾𝑔𝑔 🎀🍭https://www.patreon.com/posts/very-best-for-of-89213064 this one?

Furkan Gözükara SECourses•4/6/25, 7:01 PM

yes this is for kohya

Furkan Gözükara SECourses•4/6/25, 7:01 PM

this one onetrainer : https://www.patreon.com/posts/96028218

Patreon

OneTrainer Stable Diffusion XL (SDXL) Fine Tuning Best Presets | SE...

Get more from SECourses: FLUX, Tutorials, Guides, Resources, Training, Scripts on Patreon

Mathew•4/7/25, 3:04 AM

I saw that in 24GB_GPU_Quality_Tier_4, it's configured with "blocks_to_swap": 0,
but in 24GB_GPU_Quality_Tier_1, it's set to "blocks_to_swap": 29.

Theoretically, wouldn’t using 29 blocks to swap reduce quality/consistency?
Or is it necessary to use it, otherwise there would be a VRAM impact?

Mathew•4/7/25, 3:05 AM

I'm just getting started with the more technical parts, so I might be saying something wrong if that's the case, I apologize.

JdotBdot•4/7/25, 11:43 AM

Hi everyone! Great to be here—thanks for all the amazing tutorials and shared resources.

I’m currently running into an issue with my first Flux LoRA training on Koya. I’m on an RTX 3080 (16GB VRAM) and followed the install/setup steps as advised. All required models are downloaded. Everything seems in place, but when I try to run the training, kohya starts up, seems to run ok then without warning goes black, crashes and restarts. It’s happened several times now.

Not sure what I’m missing—or doing wrong? any advice would be hugely appreciated. Thanks in advance!

MMathew I saw that in 24GB_GPU_Quality_Tier_4, it's configured with "blocks_to_swap": 0,...

Furkan Gözükara SECourses•4/7/25, 2:38 PM

block swap only changes vram usage - thus speed not quality

JJdotBdot Hi everyone! Great to be here—thanks for all the amazing tutorials and shared re...

Furkan Gözükara SECourses•4/7/25, 2:38 PM

hello

Furkan Gözükara SECourses•4/7/25, 2:38 PM

how much RAM you have?

Furkan Gözükara SECourses•4/7/25, 2:39 PM

this sounds like out of RAM error

FFurkan Gözükara SECourses block swap only changes vram usage - thus speed not quality

Mathew•4/7/25, 3:18 PM

Oh got it, thanks a lot for clearing that up!

MMathew Oh got it, thanks a lot for clearing that up!

Furkan Gözükara SECourses•4/7/25, 3:43 PM

you are welcome

FFurkan Gözükara SECourses you are welcome

Mathew•4/7/25, 4:53 PM

Oh, just one more thing. For example, is it possible to 'pause' a LoRA at checkpoint 50 and then resume from that point to continue training up to 100 or even 200?

MMathew Oh, just one more thing. For example, is it possible to 'pause' a LoRA at checkp...

Furkan Gözükara SECourses•4/7/25, 5:02 PM

it has use last lora weights

Furkan Gözükara SECourses•4/7/25, 5:02 PM

on interface you can try

Furkan Gözükara SECourses•4/7/25, 5:03 PM

tips-info

FFurkan Gözükara SECourses it has use last lora weights

Mathew•4/7/25, 6:40 PM

Ah, got it. I'll give it a try. Thank you so much!

FFurkan Gözükara SECourses hello

JdotBdot•4/8/25, 1:59 AM

Hello, it’s an eGPU RTX3080 with 16GB RAM, task manager says it’s all available. I noticed my HDD had dropped to 50gbs free space, I’ll free more up and try again.. should I try with one of your 12gb Lora best wf’s instead?

FFurkan Gözükara SECourses this sounds like out of RAM error

JdotBdot•4/8/25, 2:01 AM

I always believed that VRAM was the same as RAM? This party always confuses me!

JJdotBdot I always believed that VRAM was the same as RAM? This party always confuses me!

Furkan Gözükara SECourses•4/8/25, 11:02 AM

totally different

Furkan Gözükara SECourses•4/8/25, 11:02 AM

RAM is system RAM and VRAM is GPU RAM

JJdotBdot Hello, it’s an eGPU RTX3080 with 16GB RAM, task manager says it’s all available....

Furkan Gözükara SECourses•4/8/25, 11:02 AM

you can show me screenshot of task manager

JJdotBdot I always believed that VRAM was the same as RAM? This party always confuses me!

404_error•4/9/25, 4:23 AM

Vram is a type of ram only for the gpu, it has way better bandwitch and optimized for certain type of data and lot of transfer. But it has a lot more latency that the ram and it too slow for random acces file and slow data that's why we got ram and vram

404_error•4/9/25, 4:25 AM

DDR5 in dual channel can transfer like 90GB/s, the GDDr7 of the 5090 can do 1792GB/s. Nothing compare

404_error•4/9/25, 2:19 PM

@Furkan Gözükara SECourses
Hello,

i run a dreambooth training on my machine and when i leave my home and come back i saw the training crash with no explication. I have a 10epoch checkpoint, (i think i was near 15), how i can resume the training ?

JdotBdot•4/9/25, 2:52 PM

Hey thanks for sharing the info on Vram , I feel that I understand it now but still not entirely sure how to measure what Vram I have! I’ve attached a screenshot from my task manager maybe someone can take a look, knowing what Vram I have will help me choose the correct Koyha Flux workflow, thanks in advance!

4404_error <@205854764540362752> Hello, i run a dreambooth training on my machine and when...

Furkan Gözükara SECourses•4/9/25, 3:47 PM

probably out of RAM

Furkan Gözükara SECourses•4/9/25, 3:47 PM

yes you can give last saved checkpoint

Furkan Gözükara SECourses•4/9/25, 3:47 PM

and it will continue - just reduce number of epochs you gonna do

JJdotBdot Hey thanks for sharing the info on Vram , I feel that I understand it now but st...

Furkan Gözükara SECourses•4/9/25, 3:47 PM

yes you have 32 GB RAM that is the issue

Furkan Gözükara SECourses•4/9/25, 3:47 PM

please upgrade to min 64 GB

Furkan Gözükara SECourses•4/9/25, 3:47 PM

more is better

JJdotBdot Hey thanks for sharing the info on Vram , I feel that I understand it now but st...

404_error•4/9/25, 8:16 PM

You have 16GB vram, it's write in "Dedicated GPU memory"

JdotBdot•4/10/25, 12:06 AM

Thanks for checking that for me! I’ll probably take the good Doc’s advice and upgrade soon — but in the meantime, out of the Lora training workflows available, which one would you recommend I try with my current setup?

JJdotBdot Thanks for checking that for me! I’ll probably take the good Doc’s advice and up...

Furkan Gözükara SECourses•4/10/25, 9:16 AM

sadly none will work very possibly but you can try this

Furkan Gözükara SECourses•4/10/25, 9:16 AM

use fp8 dev flux

Furkan Gözükara SECourses•4/10/25, 9:16 AM

use fp8 config

Furkan Gözükara SECourses•4/10/25, 9:16 AM

and select fp8 base on gradio interface and let me know if works

JdotBdot•4/11/25, 2:02 AM

Thanks Doc! Rather than struggle along I feel it will be more productive that I spend my time learning how to set up Massed Compute for Lora training for now, I’ll give it a whirl!

JJdotBdot Thanks Doc! Rather than struggle along I feel it will be more productive that I ...

Furkan Gözükara SECourses•4/11/25, 10:41 AM

great mindset

Original message was deleted

Furkan Gözükara SECourses•4/11/25, 2:51 PM

for SDXL? that can work yes

Furkan Gözükara SECourses•4/11/25, 2:51 PM

but for flux they will get mixed

Thanks! 😄 i've just upgraded my setups from 4090 to 5090 and noticing that I have problems running

Similar Threads