Software Engineering Courses (SECourses)•11mo ago

@Furkan Gözükara SECourses I’m currently training Flex on my dataset and I feel it learns much quick

@Furkan Gözükara SECourses I’m currently training Flex on my dataset and I feel it learns much quicker with half the steps compared to dev dedistilled. You were mentioning the learning rate might need adjustment. Do you think it needs to be lowered?

.tarkanOP•2/16/25, 9:00 AM

After only 4000 steps. Not even finished the whole epoch.

AiInfluence•2/16/25, 11:04 AM

@Furkan Gözükara SECourses Hey so I experimented today a bit and got lost . Can you write what a segment syntax example looks like? Like is it 2 <segment> for -1 and -2?
What would the format look like ?

AiInfluence•2/16/25, 11:04 AM

For 2 Lora faces

..tarkan @Furkan Gözükara SECourses I’m currently training Flex on my dataset and I feel ...

Furkan Gözükara SECourses•2/16/25, 11:10 AM

if learning true fast yes

Furkan Gözükara SECourses•2/16/25, 11:10 AM

lower needed

DDarth Teethius I'm trying out massed compute with your tutorial and the thindrives option says ...

Furkan Gözükara SECourses•2/16/25, 11:10 AM

yes when connecting

Furkan Gözükara SECourses•2/16/25, 11:10 AM

select end existing session

Furkan Gözükara SECourses•2/16/25, 11:10 AM

but it terminates all running apps

AAiInfluence @Furkan Gözükara SECourses Hey so I experimented today a bit and got lost . Can ...

Furkan Gözükara SECourses•2/16/25, 11:11 AM

you can ask here the dev knows better than me

https://discord.gg/5SwvsJkp

FFurkan Gözükara SECourses you can ask here the dev knows better than me 😄 https://discord.gg/5SwvsJkp

AiInfluence•2/16/25, 11:11 AM

i did didn't get an answer yet

AiInfluence•2/16/25, 11:12 AM

will it be like <segment:yolov8.pt-1,0.6.05a photo of OHWX man <Lora:Testman-1>>. ?

AAiInfluence will it be like <segment:yolov8.pt-1,0.6.05a photo of OHWX man <Lora:Testman-1>>...

Furkan Gözükara SECourses•2/16/25, 11:32 AM

that seems wrong to me

Furkan Gözükara SECourses•2/16/25, 11:34 AM

prompt like this

Furkan Gözükara SECourses•2/16/25, 11:34 AM

<segment:yolo-face_yolov9c.pt-1,0.7,0.5>photograph of ohwx man

Furkan Gözükara SECourses•2/16/25, 11:34 AM

and then click lora

AiInfluence•2/16/25, 11:34 AM

and for Lora?

AiInfluence•2/16/25, 11:34 AM

oh ok

Furkan Gözükara SECourses•2/16/25, 11:34 AM

but actually i dont know how to use lora

Furkan Gözükara SECourses•2/16/25, 11:34 AM

on prompt

Furkan Gözükara SECourses•2/16/25, 11:34 AM

but it should come at the end

Furkan Gözükara SECourses•2/16/25, 11:34 AM

appended

FFurkan Gözükara SECourses but actually i dont know how to use lora

AiInfluence•2/16/25, 11:34 AM

it say in the document after the text prompt to type manually<Lora...>

AiInfluence•2/16/25, 11:35 AM

<segment:yolo-face_yolov9c.pt-1,0.7,0.5>photograph of ohwx man <Lora....>

AiInfluence•2/16/25, 11:35 AM

yeah I think that's jt thanks doc !

AAiInfluence yeah I think that's jt thanks doc !

Furkan Gözükara SECourses•2/16/25, 11:37 AM

FFurkan Gözükara SECourses RTX 5090 faster than L40S with FLUX at BF16

Timson•2/16/25, 12:29 PM

Do you meat at training? At inference L40S is only 8% slower but 0.3 second new prompt delay opposed to >3s on 5090

Timson•2/16/25, 12:31 PM

so for practical purposes L40S is still better

TTimson Do you meat at training? At inference L40S is only 8% slower but 0.3 second new ...

Furkan Gözükara SECourses•2/16/25, 1:00 PM

i think it depends if you are constantly changing prompt yes

Furkan Gözükara SECourses•2/16/25, 1:00 PM

but L40S is 8000+ USD

Furkan Gözükara SECourses•2/16/25, 1:01 PM

and 5090 is at its slowest performance

Furkan Gözükara SECourses•2/16/25, 1:01 PM

it will only get better we dont even have official pytorch for windows

phoenix•2/16/25, 1:34 PM

@Furkan Gözükara SECourses Hello I just started training with the 24 GB flux dev model but my storage was less only 50gb so i think it can create only 2 checkpoints but I checked the training status and the second checkpoint epoch has already strated even before checkpoint one epochs are totally complete.How to save one checkpoint to huggingface if the second one has also started simultaneously

phoenix•2/16/25, 1:34 PM

@Furkan Gözükara SECourses and also once the storage is full what error will I get and how to resume the training

phoenix•2/16/25, 1:35 PM

@Furkan Gözükara SECourses actually I have set 4 checkpoints instead of 2

Pphoenix @Furkan Gözükara SECourses Hello I just started training with the 24 GB flux dev...

Furkan Gözükara SECourses•2/16/25, 1:37 PM

i never could make it save to hugging face before automatically from kohya

Furkan Gözükara SECourses•2/16/25, 1:37 PM

once storage is full i dont remember behaviour sadly

phoenix•2/16/25, 1:42 PM

@Furkan Gözükara SECourses So can i just stop this process and increase the storage for the same gpu instance and the again start training

Pphoenix @Furkan Gözükara SECourses So can i just stop this process and increase the stor...