ah its 77 for CLIP, 512 for T5
ah its 77 for CLIP, 512 for T5
Fyep
Fover batch 2
Fdoesnt bring any speed
Fbatch 2 brings slight like 20%
Fhello this is internet error
Fyou can fix with warp vpn - cloudflrare
FInstall the free app that makes your phone’s Internet more fast, private, and reliable.
Fso activate warp vpn and do a fresh install
OYes... but fp8 just not working with Lora. I don't know why. I just switch models, everything else stay the same.
Maybe I'll risk a forge update
Maybe I'll risk a forge update
LThe problem is not with the internet, I installed the original one and it works https://github.com/bmaltais/kohya_ss and your build is failing
GitHub
Contribute to bmaltais/kohya_ss development by creating an account on GitHub.
FYou can see my install by editing file it is nothing special it installs that with accurate branch 
FAh could be forge thing
FI used only Fp16 on forge
SJust a quick FYI for anyone interested, here's how long it took me to do a 20img/150 epoch training session on a 4090 this morning. Used 'Rank 4 best config better colours' script
FVery nice speed
LIs this error going to cause me problems down the road?
LThanks
FNope
FWe don't use xformers with FLUX
OThe fp16 works really well.
I used no captioning, no regs. Best Lora i ever made
I used no captioning, no regs. Best Lora i ever made
D@Dr. Furkan Gözükara Can you please tell me if this is considered a good learning curve for my 3090?
Floss rate is meaningless
Fspeed is looking ok not very bad
Fi think your model will be really good
Ftrained a lora for a client i think it turned out to be perfect
Fcan do smiling very well since we included that :d
MWhile you guys are training Flux, I'm experiencing the interesting features of PonyRealism 2.1 under Prodigy. (I see that the Flux training is still being tweaked at Kohya, so I'll wait.) It turns out that the SDXL rule of only training Unet doesn't apply here, because then the finished Lora will only look a little like the person. However, if I train normally (i.e. without the --network_train_unet_only switch), I'm already on target. By using the 'fp8 base training' switch I can weaken the strength of the model a bit, i.e. sometimes it can come in handy in achieving a more accurate Lora. Without fp8, the Lora is stronger, so the d_coef switch needs to be reduced.
Ffor sdxl i suggest unet + text encoder 1
Falso for flux i am testing training clip large text encoder atm
DRealVisXL5 was released
Fwow i should test
Ftesting flux clip large trainings
Dseems to be a big training effort, not just remerge: "Hello everyone! An early version of RealVisXL V5.0 (560k steps) is now available on Hugging Face. The model is still in training, and the final version will be released on Hugging Face and CivitAI at the end of August. You can leave your feedback on this version in the comments below."
G@Dr. Furkan Gözükara Bro, I am going to sleep now its 6am haha, training loras all night, I fixed all my machines now. Anyway, I want to Train 1000 images (a style of art) I want to train that but doing it on 1 A6000 it would take like a week right? What system you recommend and how much would it cost me on Mass compute? 4-8 GPU A6000? If its $3-5 an HR I am down to pay it. Please do recommend the correct VM to use on there and do you have a video setting one up as I only use local, never rented hardware before. Let me know thanks and gn brother! Totally worth the Patreon fee!
Fget 4x A6000 on massed compute
Fit is 1.25$ per hour with our coupon
Fi made the multi gpu training tutorial so it will be published today or tomorrow hopefully
V@Dr. Furkan Gözükara two models have been changed by the author, %cd /kaggle/temp/VAE
!wget "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/ae.safetensors" -O "ae.safetensors" !wget "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/flux1-schnell.safetensors" -O "flux1-schnell.safetensors" am I right? another question it seems to take a few mins to get 1 image from flux model on kaggle, right?
!wget "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/ae.safetensors" -O "ae.safetensors" !wget "https://huggingface.co/black-forest-labs/FLUX.1-schnell/resolve/main/flux1-schnell.safetensors" -O "flux1-schnell.safetensors" am I right? another question it seems to take a few mins to get 1 image from flux model on kaggle, right?
V@Dr. Furkan Gözükara another question, if I want to use custom nodes in swarm UI, I did not see the folder, do I have to create one? Autocompletions Users-log.ldb clip_vision tensorrt
Embeddings Users.ldb comfy-auto-model.yaml unet
Lora VAE controlnet upscale_models
Stable-Diffusion clip diffusion_models, or where are the custom node folder?
Embeddings Users.ldb comfy-auto-model.yaml unet
Lora VAE controlnet upscale_models
Stable-Diffusion clip diffusion_models, or where are the custom node folder?
M@Dr. Furkan Gözükara I've test on windows on my PC with A4000 (8go) and it worked, however as I suspected it use shared memory during the step "move text encoder to gpu" as there is not enough memory on the GPU, that's why it OOM on linux, because it can't use shared memory, then it's passed back to the cpu freeing memory and the rest is loaded that require less than 8go so during training it's stay under 8go. Is there a way to not move the text encoder to gpu, like having it calculated on cpu directly, having it use shared memory or making it use less VRAM, so it will work also on linux ?
Fyes links changed
Fi see what you mean now
Fso you get error during cache
Fyou need to ask this to kohya
Fcant linux has shared vram?
MOk I could maybe create a ticket on it's git
MI search but didn't find a way to do that it seem not implemented in linux nvidia driver
Fplease make a detailed reply here and tag @kohya-ss : https://github.com/kohya-ss/sd-scripts/pull/1374
GitHub
Replace SD3Tokenizer with the original CLIP-L/G/T5 tokenizers.
Extend the max token length to 256 for T5XXL.
Refactor caching for latents.
Refactor caching for Text Encoder outputs
Extract arch...
Extend the max token length to 256 for T5XXL.
Refactor caching for latents.
Refactor caching for Text Encoder outputs
Extract arch...
