Just a quick LoRA I pulled together to test the training capabilities of my GPU on Qwen.
Description
FAQ
Comments (6)
How did you train it? It seems to be at its best when using FP8 (non-distilled) in combination with the Lightning Lora @ 8steps.
I used Ostris's AI-toolkit with an RTX 3090. Had to quantize the model and text encoder to 4-bit though
LususNaturae Hey man can you share your settings? I want to train a Sinozick style LoRA (X.com/sinozick), I wonder if I can do it with my 4060 Ti with 16 of vram, I'm interested in training style loras, I've trained some for Illustrious in a about 2 to 3 hours
LususNaturae Btw the lora isn't to steal the artist style, but to use it mixed with other styles etc haha
Henwo But keep in mind my config was using around 22 GiB of VRAM
https://pastebin.com/vVnBXmJV
LususNaturae THANKS! I see, let's see if I can run it haha, maybe it offloads to system ram