Software Engineering Courses (SECourses)•15mo ago

I'm looking to do a finetune with a fp8 but the Dr. setting for it uses shared memory, and is slow a

I'm looking to do a finetune with a fp8 but the Dr. setting for it uses shared memory, and is slow at around 10s/it. But speaking to koyha, it seems its an intended behavior to use shared memory.

VVigilence I'm looking to do a finetune with a fp8 but the Dr. setting for it uses shared m...

wardensc2•10/9/24, 4:49 PM

just learn to use masscompute service or runpod. Local VGA is not mean to change model anymore. Quality is worst.

VigilenceOP•10/9/24, 4:49 PM

I might have to, but I can do super fast loras, so I was suprised how slow it was.

wardensc2•10/9/24, 4:50 PM

I use 3090 with 24gb Vram but I can only train fp8 lora. I use mass compute to train fully fp16 lora with the cost of about 4 usd

wardensc2•10/9/24, 4:50 PM

very cheap

VigilenceOP•10/9/24, 4:50 PM

That is per hour?

wardensc2•10/9/24, 4:50 PM

no, total cost

OinkSauce•10/9/24, 4:51 PM

Massed Compute - from my experience has been great, far better than runpod. The SECourses pricing on the A6000's is great.

wardensc2•10/9/24, 4:51 PM

per hour is only 1.25 usd

wardensc2•10/9/24, 4:52 PM

when training with my local pc, I need 11 hours of 3090 run full throttle, my poor pc fan roar the hell of of my room and the temperature turn up a lot

wardensc2•10/9/24, 4:52 PM

so 4 usd saving my trouble and time

wardensc2•10/9/24, 4:53 PM

11 hours of 3090 for fp8 lora

wardensc2•10/9/24, 4:53 PM

no thank you

OOinkSauce If anyone has a second... Could i get more information on what "de-distilled" m...

dsienra•10/9/24, 4:53 PM

I'm training this model now, Works much better for training, you can train multiple people on one lora without concept mixing, it behaves very much like sdxl, much more flexible

VigilenceOP•10/9/24, 4:53 PM

Yeah the pc lights stayed on and it looks like a disco ballroom.

Ddsienra I'm training this model now, Works much better for training, you can train multi...

wardensc2•10/9/24, 4:54 PM

I never see any loras in SDXL have multiple concept, Flux maybe the first model can do it

dsienra•10/9/24, 4:55 PM

I trained loras on sdxl with more than 20 people no problem

wardensc2•10/9/24, 4:56 PM

Since SDXL have instance ID, so I use it for more convenience if I need people lora anyway

VigilenceOP•10/9/24, 4:57 PM

Best gpu prices are from here - https://www.autodl.com/home but I don't know how to run them. Learned about them via an article discussing 48gb 4090s.

AutoDL算力云 | 弹性、好用、省钱。租GPU就上AutoDL

AutoDL为您提供专业的GPU租用服务，秒级计费、稳定好用，高规格机房，7x24小时服务。更有算法复现社区，一键复现算法。

Ddsienra I trained loras on sdxl with more than 20 people no problem

wardensc2•10/9/24, 4:57 PM

this really new to me

Ddsienra I'm training this model now, Works much better for training, you can train multi...

OinkSauce•10/9/24, 4:57 PM

Is it possible to create a finetune, then extract the LORA? I've had good luck with that for single subjects, and sometimes it's also nice to have a checkpoint as well.

dsienra•10/9/24, 4:57 PM

flux-dev can train only one subject, but with flux-dev-de-distill I can learn many concepts

VVigilence Best gpu prices are from here - https://www.autodl.com/home but I don't know ho...

wardensc2•10/9/24, 4:58 PM

Most of us dont know chinese, oh well

VigilenceOP•10/9/24, 4:58 PM

@wardensc2 I just use google translate to change the text.

OOinkSauce Is it possible to create a finetune, then extract the LORA? I've had good luck ...

dsienra•10/9/24, 4:59 PM

yes but is too slow... now I'm training on my 3090ti, may be then I will try on mass compute

OinkSauce•10/9/24, 4:59 PM

I don't think I've ever tried to train a finetune with more than one subject

VVigilence @wardensc2 I just use google translate to change the text.

wardensc2•10/9/24, 4:59 PM

but when you access to their server, their GUI also chinese, what will you do = )))

VigilenceOP•10/9/24, 5:00 PM

@wardensc2 Click translate in the browser, lol.

VigilenceOP•10/9/24, 5:00 PM

Wwardensc2 Since SDXL have instance ID, so I use it for more convenience if I need people l...

dsienra•10/9/24, 5:00 PM

yes but the quality is much worse... now I'm full on flux-dev-de-distill

wardensc2•10/9/24, 5:01 PM

Also @Dr. Furkan Gözükara already set up kohya_ss for quick install on Mass Compute, I just need 1 console code for it to run everything

VVigilence Click to see attachment

wardensc2•10/9/24, 5:01 PM

no this is only the beginning

wardensc2•10/9/24, 5:02 PM

similar to masscompute, you have to upload your image training data and the checkpoint you want to train, vae, clip and install Kohya as well

wardensc2•10/9/24, 5:02 PM

it similar to our local PC but better hardware, it's not one click training.

Ddsienra yes but the quality is much worse... now I'm full on flux-dev-de-distill

wardensc2•10/9/24, 5:04 PM

No I guarantee when it come to people image, install ID is much better than lora if you know how to use it properly

wardensc2•10/9/24, 5:04 PM

I'm not sure I can upload AI image of my instant ID for you to see on discord

wardensc2•10/9/24, 5:05 PM

I make Mbappe Instant ID, not sure SDXL lora can beat it

Wwardensc2 No I guarantee when it come to people image, install ID is much better than lora...

dsienra•10/9/24, 5:05 PM

I know what you mean I try it, but the quality over all is much better on flux

dsienra•10/9/24, 5:06 PM

SDXL is good but not as good

wardensc2•10/9/24, 5:06 PM

of course I dont compare with Flux but SDXL lora can't beat Instant ID

VigilenceOP•10/9/24, 5:06 PM

@wardensc2 I understand. Im just pointing at the price difference.

NVIDIA L20 / 48GB single precision 59.35 TFLOPS / semi precision 119.5 Tensor TFLOPS ￥ 4.19 / hour

USD is 0.60 an hour.

VigilenceOP•10/9/24, 5:07 PM

NVIDIA L40 / 48GB single precision 90.5 TFLOPS / semi precision 181.05 Tensor TFLOPS ￥ 4.82 / hour

$0.69 USD

Wwardensc2 of course I dont compare with Flux but SDXL lora can't beat Instant ID

dsienra•10/9/24, 5:08 PM

IDK with a well trained lora you get perfect likeness and more flexibility

VVigilence @wardensc2 I understand. Im just pointing at the price difference. NVIDIA L2...

wardensc2•10/9/24, 5:08 PM

yes I know but when you using it, you can't translate everything if the operating system and other support software are also chinese

VigilenceOP•10/9/24, 5:08 PM

I see, your saying the whole OS would be in chinese.

dsienra•10/9/24, 5:08 PM

but its ok if works for your purpose

VigilenceOP•10/9/24, 5:09 PM

Maybe Dr. Can team up with them and have them make an option for flux training preconfigured.

Ddsienra IDK with a well trained lora you get perfect likeness and more flexibility

wardensc2•10/9/24, 5:09 PM

normally I just gen a image with style and pose I want then apply Instant ID on it. SDXL lora can't flexible than this

wardensc2•10/9/24, 5:10 PM

something like this

dsienra•10/9/24, 5:10 PM

looks very good

wardensc2•10/9/24, 5:11 PM

this is SDXL not Flux

I'm looking to do a finetune with a fp8 but the Dr. setting for it uses shared memory, and is slow a

Similar Threads