Testing "Forge with CUDA 12.1 + Pytorch 2.3.1" and "Forge with CUDA 12.4 + Pytorch 2.4" to see what
Testing "Forge with CUDA 12.1 + Pytorch 2.3.1" and "Forge with CUDA 12.4 + Pytorch 2.4" to see what the difference are
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=16)
Fye i hate this c++ thingie
Fit is time to finalize it :d
![Dazzastrous [4090]](https://cdn.discordapp.com/avatars/871389071686135838/c4a5f4d81d9eeb18d856295d692e2c07.webp?size=40)
Dyeh me too
Fi will make a video for this
KDamn
D30.2 secs
KYeah well... 4090 sounds right
Ffaster internet so useful
Fswarmui has --fast
Ffor rtx 4090
Freduces flux to like 10-12 second
F20 steps
D
Doink oink
Dflux is a bit heavy on the boka , SD more detailed backgrounds
Dmore detail with sd
D
Dlike the colours in this
Fhaha i made a thread on all nvidia githubs
Fhttps://github.com/NVIDIA/CUDALibrarySamples/issues/217
https://github.com/NVIDIA/cccl/issues/2383
https://github.com/NVIDIA/cutlass/issues/1793
https://github.com/NVIDIA/cuda-samples/issues/297
https://github.com/NVIDIA/TensorRT/issues/4114
https://github.com/NVIDIA/cuda-python/issues/88
https://github.com/NVIDIA/cudnn-frontend/issues/106
lol they have 500 repos
https://github.com/NVIDIA/cccl/issues/2383
https://github.com/NVIDIA/cutlass/issues/1793
https://github.com/NVIDIA/cuda-samples/issues/297
https://github.com/NVIDIA/TensorRT/issues/4114
https://github.com/NVIDIA/cuda-python/issues/88
https://github.com/NVIDIA/cudnn-frontend/issues/106
lol they have 500 repos
FPatreon
Get more from SECourses: Tutorials, Guides, Resources, Training, FLUX, MidJourney, Voice Clone, TTS, ChatGPT, GPT, LLM, Scripts by Furkan Gözü on Patreon
Fi will update article testing still
SThe worst thing about running the kohya flux training update. "Haha you fewl, you installed torch 2.5. I will now downgrade you to 2.4 and complain about dependencies needing 2.5" And then the installer runs off like a 1920's villain and ties a damsel to train tracks
Fi overwrite gui file to prevent it
to prevent it
Xwill give this a try now
D@Dr. Furkan Gözükara One question, I'm trying to train flux with a big dataset of 1500 images, with the sugested learning rate is too slow to converge and get any results, I'm trying now with 4e-4 lr and I getting results much faster and until now looks good, do you think is too aggressive lr? may be I can start with this high lr and then finalize the training whith a lower lr to learn more finer details but without having to wait 15 days of training
Fwell the LR bigger certainly it will learn faster but it is like you are slicing the cake with a bigger knife so the minor detailed cuts becomes impossible
Fyou can speed up training with multiple GPU
Falso bigger batch size requires bigger LR
Dyou right, but by now I doing everything on my local computer, do you think that finalize the training with a lower lr can improve the final results, startin with a high learning rate and when is aready on track I can reduce the lr, do you think that lowerimg the lr at the end can be beneficial?
Xjust started training 
FMan I need 48 gb rtx 5090 comon Nvidia 
FI also need a mb that can perfectly run 3 gpu at the same time so I can test all 
F3060 3090 and 5090
XThe way you explain everything in this video so far is just amazing Dr i really missed learning from you
SJust a quick FYI on flux training. Using installer v30 on an i9-13900 with 64gig ram and a 4090. Best Config Rank 4 quoted as 4.85 secs/IT, I'm getting 3.60/IT. So even quicker than expected
Sbasically 1000 IT per hour
Fthank you so much
F4090 is a beast
Fi am testing rope pearl face fusion with new c++ compiler
F
Fcuda 11.8 success
CFor SDXL, does anyone know why using a lightning lora makes use of my person lora better?
The one on the right is with lightning lora, which actually looks better!
Is there something going on with my training parameters that causes this issue?
Trained on 64/1 network dims with LR of 0.0006. but why would lightning lora cause it to be better???
The likeness is actually better as well, so its not like its ignoring my lora
The one on the right is with lightning lora, which actually looks better!
Is there something going on with my training parameters that causes this issue?
Trained on 64/1 network dims with LR of 0.0006. but why would lightning lora cause it to be better???
The likeness is actually better as well, so its not like its ignoring my lora
XI have RTX 4090 and getting
Steps: 0% | 10/3000 721.04s/it … it’s been around 2 hours already, is this normal?
Steps: 0% | 10/3000 721.04s/it … it’s been around 2 hours already, is this normal?
X@Furkan Gözükara SECourses how long does training usually take?
Fwhich gpu you have?
