Software Engineering Courses (SECourses)•10mo ago

What is the easiest way to train a LoRA locally / and not locally? 😄

What is the easiest way to train a LoRA locally / and not locally?

Ccilium What is the easiest way to train a LoRA locally / and not locally? 😄

Furkan Gözükara SECourses•3/27/25, 12:53 PM

kohya

Furkan Gözükara SECourses•3/27/25, 12:53 PM

with our config

Furkan Gözükara SECourses•3/27/25, 12:53 PM

what gpu you have?

ciliumOP•3/27/25, 1:07 PM

NVIDIA 4070

Ccilium NVIDIA 4070

Furkan Gözükara SECourses•3/27/25, 1:17 PM

it is a good one. if you have 64 gb ram you can train easily

ciliumOP•3/27/25, 1:17 PM

i have 32, does the more RAM u have make it so u can train faster too? So its a mix of Video card and RAM?

Ccilium i have 32, does the more RAM u have make it so u can train faster too? So its a ...

Furkan Gözükara SECourses•3/27/25, 1:18 PM

without 64 gb ram sadly it fails

Furkan Gözükara SECourses•3/27/25, 1:19 PM

because we do block swapping

Furkan Gözükara SECourses•3/27/25, 1:19 PM

so when training it moves non trained blocks to ram

Furkan Gözükara SECourses•3/27/25, 1:19 PM

and get from ram going to get trained blocks

Furkan Gözükara SECourses•3/27/25, 1:19 PM

this way the model fits into gpu with minimal speed loss

HHiramos_Guys, what are your Wan2.1 img-to-vid 480 14B generation times like after the la...

J2_Nedo•3/27/25, 1:28 PM

what resolution are you using for the videos?

FFurkan Gözükara SECourses without 64 gb ram sadly it fails

Weatherby•3/27/25, 4:10 PM

Does cuddims at like, 8400 help this at all, or is it just more ram regardless of speed is better?

Weatherby•3/27/25, 4:11 PM

(Arrow lake doesn't like more than 2 dimms without memory issues. Went from 128 to 48.. but 6000 running at 5400 to 8400)

ciliumOP•3/27/25, 6:37 PM

Are you guys playing with the new Chat GPT image generation

ciliumOP•3/27/25, 6:37 PM

it seems kind of silly good

WWeatherby Does cuddims at like, 8400 help this at all, or is it just more ram regardless o...

Furkan Gözükara SECourses•3/27/25, 6:39 PM

well it makes little difference

Ccilium Are you guys playing with the new Chat GPT image generation

Furkan Gözükara SECourses•3/27/25, 6:39 PM

they will probably cripple it soon :d

ciliumOP•3/27/25, 6:41 PM

Yeah they already announced 2 hours ago but the subtle changes it can make just from text, seem quite good

Ccilium Yeah they already announced 2 hours ago but the subtle changes it can make just ...

Furkan Gözükara SECourses•3/27/25, 6:49 PM

yes its text feature looking really good

Furkan Gözükara SECourses•3/27/25, 6:49 PM

i am expecting local similar model :d

ciliumOP•3/27/25, 7:04 PM

Local damn

ciliumOP•3/27/25, 7:04 PM

that would be a complete gamechanger

ciliumOP•3/27/25, 7:04 PM

What is a good way to just simply get a good LoRA trained quickly by paying? like whatever it costs to get it fast and izi

Inoshi•3/27/25, 7:09 PM

4o model is crazy you can pretty much do any pose/packground/ character/ style and then move it to a good inpaint/depth/lora combination to get what you want.. but i cant wait until Open Source community catches up to them. maybe until the end of the year?

Ccilium What is a good way to just simply get a good LoRA trained quickly by paying? lik...

Furkan Gözükara SECourses•3/27/25, 7:16 PM

get a good gpu and train with our configs

IInoshi 4o model is crazy you can pretty much do any pose/packground/ character/ style a...

Furkan Gözükara SECourses•3/27/25, 7:16 PM

100% i expect

IInoshi 4o model is crazy you can pretty much do any pose/packground/ character/ style a...

ciliumOP•3/27/25, 8:39 PM

it seems like even LoRAs might not be needed soon

ciliumOP•3/27/25, 8:39 PM

My friend just made a character and used him in future generations and its almost perfectly consistent

ciliumOP•3/27/25, 8:40 PM

But isn't this expensive for OpenAI and Google AI studio etc, wouldn't they have to just gate it over a bigger paywall like 200$ or something?

ciliumOP•3/27/25, 8:40 PM

everyone generating infinite for 20$

Inoshi•3/27/25, 8:40 PM

takes too long now in comparison with dalle 3

Inoshi•3/27/25, 8:41 PM

havent tested sora since 4o image announcement to check if the waiting duration got bigger

IInoshi takes too long now in comparison with dalle 3

ciliumOP•3/27/25, 8:41 PM

But that's because 1 trillion ppl are using it

Inoshi•3/27/25, 8:42 PM

i think the biggest breathrough in LLM was the chinese one this year where they showed the can have strong models way smaller in size

Inoshi•3/27/25, 8:42 PM

now multimodality open source models are entering which is great

ciliumOP•3/27/25, 8:43 PM

People keep saying this and then this image generation model can tweak almost anything with simple text, seems for the average joe its pretty much the main thing they need

Inoshi•3/27/25, 8:44 PM

i think this is going to actually end some starting careers in graphic designing, they will keep the pros the rest can move on

⚔

⚔Seifme⚔•3/28/25, 9:36 AM

Hey everyone, I have dual RTX 3090 linked with NVLINK, I've tried several ways to train lora using multi GPUs but I always fail & get errors, I'm on windows & I tried Kohya SS & Ai-toolkit with and without accelerate launch with no luck, I'm not a developer & I don't know why I get an error everytime

So if anyone can help me using the installation

Zono50•3/28/25, 12:00 PM

@Furkan Gözükara SECourses how do i run cog2vlm in linux and have it use fp8?

Inoshi•3/28/25, 12:19 PM

https://github.com/FoundationVision/VAR

GitHub

GitHub - FoundationVision/VAR: [NeurIPS 2024 Best Paper][GPT beats ...

[NeurIPS 2024 Best Paper][GPT beats diffusion

] [scaling laws in visual generation

] Official impl. of "Visual Autoregressive Modeling: Scalable Image Generation via Next-Scale Prediction&...

GitHub - FoundationVision/VAR: [NeurIPS 2024 Best Paper][GPT beats ...

IInoshi https://github.com/FoundationVision/VAR

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•3/28/25, 12:25 PM

How do you find these repositories?

Inoshi•3/28/25, 1:04 PM

i am into ai but also people i follow post these regularly, and i share those i believe can have an impact based on the recent trend (4o)to places i believe someone might be interested