Furkan Gözükara
patreon
How to Convert Your Trained FLUX, Stable Diffusion 3.5, SDXL and Other LoRAs Into FP8 and FP16 to Save Huge Space With Kohya GUI
🕑 Added 2024-12-10 10:01:00 +0000 UTCPatreon exclusive posts index to find our scripts easily, Patreon scripts updates history to see which updates arrived to which scripts and amazing Patreon special generative scripts list that you can use in any of your task.
Join discord to get help, chat, discuss and also tell me your discord username to get your special rank : SECourses Discord
Please also Star, Watch and Fork our Stable Diffusion & Generative AI GitHub repository and join our Reddit subreddit and follow me on LinkedIn (my real profile)
=======
10 December 2024 Update
Screenshot updated according to latest version
Details
As you know I have finalized and perfected my FLUX LoRA Training workflow until something new arrives
Configs and necessary explanation are shared here : https://www.patreon.com/posts/110879657
LoRA full tutorial : https://youtu.be/nySGu12Y05k
Fine Tuning / DreamBooth full tutorial : https://youtu.be/FvpWy1x5etM
LoRA Extract from Fine-Tuned / DreamBooth checkpoint : https://www.patreon.com/posts/112335162
How To Reduce LoRA Model Files Size
There are 3 ways of reducing generated / save LoRA models
1 : When training, you can reduce LoRA Network Rank (Dimension)
I don't recommend this since it reduces the quality
2 : Saving as FP16 when training
This makes it half but still files will be big, almost no quality loss as shown in below grid
3 : Saving as FP8 after trained as shown in above images via Kohya GUI
As shown in below grid almost best quality and 2.5 GB FP32 LoRA file becomes 634 MB
Full grid that shows FP8 vs FP16 vs FP32 : LoRA_Precision_FP8_vs_FP16_vs_FP32_Grid.jpg
With this way you can convert any model LoRAs like:
Stable Diffusion 1.5 (SD 1.5), Stable Diffusion XL (SDXL), Stable Diffusion 3 or Stable Diffusion 3.5 (SD 3.5), FLUX and such
However how other architectures like SDXL will perform need to be grid tested
Comments
Furkan Gözükara
100% related to your windows setup. i can connect your pc and fix if you upgrade to gold tier. or look for drive permissions on google
Emelaine Pilan
impossible to convert... always the same permission error, serialize_file(_flatten(tensors), filename, metadata=metadata) safetensors_rust.SafetensorError: Error while serializing: IoError(Os { code: 5, kind: PermissionDenied, message: "Acesso negado." })
Furkan Gözükara
no atm but i am pretty sure bmaltais can implement can you open request here? : https://github.com/bmaltais/kohya_ss/issues - i can also reply
sebastian
Is there a way to batch do all the saved steps along the way?
Furkan Gözükara
Kohya got SD 3.5 training hopefully I will research this week
cool1
is says "or stable diffusion 3.5"... is it true that kohya hasn't been updated for training SD3.5 large yet? Is onetrainer the only thing that does (I haven't used ontrainer).
Furkan Gözükara
it is super fast. you can pick gpu too :)
Yassine Belatar
Why cpu not gpu isn't cpu slower ? Also How long it takes ?
