Software Engineering Courses (SECourses)•10mo ago

we were using fp32 weights

Furkan Gözükara SECoursesOP•3/4/25, 4:25 PM

now we will use fp16 half space

Furkan Gözükara SECoursesOP•3/4/25, 4:25 PM

but you can set precision to fp32 in app which would double VRAM

Ssergio_gonzo Dr Furkan this graphic show thta the 5070 ti have AI TOPS Equivalent to 4090 th...

Furkan Gözükara SECoursesOP•3/4/25, 4:26 PM

nope

Furkan Gözükara SECoursesOP•3/4/25, 4:26 PM

it has lesser VRAM

Furkan Gözükara SECoursesOP•3/4/25, 4:26 PM

pay attention to cuda cores that is what matters most + vram

sergio_gonzo•3/4/25, 4:26 PM

imean 4070 ti super, vs 5070 ti super in AI task thediference of AITOPS ahve any impatc between them?

FFurkan Gözükara SECourses we were using fp32 weights

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•3/4/25, 4:28 PM

So your app was already using those 6 parts of the model and it's fp32 rught? And we are able to use those 6 part foles through your app on 48gb vram?

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 So your app was already using those 6 parts of the model and it's fp32 rught? An...

Furkan Gözükara SECoursesOP•3/4/25, 4:29 PM

it was using fp32 weights in bf16

Ssergio_gonzo imean 4070 ti super, vs 5070 ti super in AI task thediference of AITOPS ahve an...

Furkan Gözükara SECoursesOP•3/4/25, 4:29 PM

i expect 10% difference between 2

FFurkan Gözükara SECourses it was using fp32 weights in bf16

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•3/4/25, 4:29 PM

Can we now use those weights in fp32 through your app?

�𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮 Can we now use those weights in fp32 through your app?

Furkan Gözükara SECoursesOP•3/4/25, 4:30 PM

yes both works

Furkan Gözükara SECoursesOP•3/4/25, 4:30 PM

just dont download new model weights :d

FFurkan Gözükara SECourses yes both works

�

𝗡𝗶𝘁𝗶𝗻 𝗣𝗮𝗿𝗲𝘃𝗮•3/4/25, 4:31 PM

That's great so what changes we need to select in the ui in order to use those weights in fp32

Furkan Gözükara SECoursesOP•3/4/25, 4:36 PM

change torch_dtype

J2_Nedo•3/4/25, 4:43 PM

so instead fp16 using fp32 this will be much slower right? i have to test to see if there is a difference in the video

JJ2_Nedo so instead fp16 using fp32 this will be much slower right? i have to test to see...

Furkan Gözükara SECoursesOP•3/4/25, 4:59 PM

i tested fp16 weight vs fp32 weight

Furkan Gözükara SECoursesOP•3/4/25, 4:59 PM

no difference

Furkan Gözükara SECoursesOP•3/4/25, 4:59 PM

but inference in fp32 i didnt test as it would require

Furkan Gözükara SECoursesOP•3/4/25, 4:59 PM

astronomical vram

sergio_gonzo•3/4/25, 5:00 PM

Dr Furkan this model WAN could work in AMD gpus?

Ssergio_gonzo Dr Furkan this model WAN could work in AMD gpus?

Furkan Gözükara SECoursesOP•3/4/25, 5:01 PM

nope

Furkan Gözükara SECoursesOP•3/4/25, 5:01 PM

i mean at least my code

Furkan Gözükara SECoursesOP•3/4/25, 5:01 PM

amd has to bring like 64 gb consumer gpus

Furkan Gözükara SECoursesOP•3/4/25, 5:01 PM

then community will make it work

Furkan Gözükara SECoursesOP•3/4/25, 5:01 PM

but for mere 24 gb no one gonna waste time

Jjulius @Dr. Furkan Gözükara Is it possible to control which face gets the inpaint if th...

julius•3/4/25, 5:02 PM

or anyone else knows how to fix this?

sergio_gonzo•3/4/25, 5:03 PM

the incremental in amount of vram remember me when intel only give us 4 cores in mainstrain,,, AMD must do somthing to change this

Jjulius @Dr. Furkan Gözükara Is it possible to control which face gets the inpaint if th...

julius•3/4/25, 5:14 PM

found this in the docs. You can control the index apparently

Jjulius found this in the docs. You can control the index apparently

julius•3/4/25, 5:16 PM

Also seems like you can check the segmentation mask here https://github.com/mcmonkeyprojects/SwarmUI/issues/7#issuecomment-2187549309

GitHub

Yolo segmentation seems to only detect one face · Issue #7 · mcmonk...

Hi, First, thank you for this awesome UI! I'm trying to generate an image, where all faces will be fixed, like in the ADetailer extension of A1111. Here is my prompt: oil painting of a male act...