GTX 1660 6G always shows "step: loss=nan, smooth loss=nan". Help me solve this problem

[Eger1983]

Hello. I am creating empty LORA. Video card GTX 1660. I don’t understand what needs to be configured.

[mx]

You have not posted enough information here (you need the entire config json, not just the single screenshot, as well as details about your dataset).

[Eger1983]

I once managed to create low-quality LORA, but the settings were not saved. Current settings are in the attachment. The dataset is as follows: photo 1024x1024 (tried from 5 to 15 photos) description in .txt.
config_test.json

[mx]

Some things I've noticed:

(1) You've repeated the same concept like 10 separate times.
(2) You are training SDXL but have the resolution set at 320. I do not think that will work. SDXL's native resolution is 1024.
(3) I doubt you could train SDXL on 6GB VRAM anyway.

[Heasterian]

16 series GPU's have poor support of FP16. Try to train FP32 with autocast cache disabled (and recreate cache ofc).

[Heasterian]

You most likely will need to use ADAFACTOR optimizer instead of ADAMW to not get OOM.

[Eger1983]

Thank you. I will try.