using the latest sd3 branch and trying to train a FLUX1.dev LoRA using AdamW and a cosine or linear scheduler produces the error. See the attached config for complete details.
I believe starting from this change there has been a problem introduced with the cosine and linear schedulers. This is true at least for FLUX dev1 LoRA training using torch 2.4.0:
https://github.com/kohya-ss/sd-scripts/pull/1393
This is the config TOML that produces the error for me:
config_lora.zip
This is in reference to: bmaltais/kohya_ss#2812
using the latest sd3 branch and trying to train a FLUX1.dev LoRA using AdamW and a cosine or linear scheduler produces the error. See the attached config for complete details.
I believe starting from this change there has been a problem introduced with the cosine and linear schedulers. This is true at least for FLUX dev1 LoRA training using torch 2.4.0: https://github.com/kohya-ss/sd-scripts/pull/1393
This is the config TOML that produces the error for me: config_lora.zip