What is the reason that the trained LoRA will only splash screen?

[yetone]

Hello everyone, I am new to this, I would like to ask why my trained LoRA only splashes, I have trained with different image sets many times with this result, is it because of the CUDA version problem? My CUDA is version 12.1

[kohya-ss]

I'm not sure the CUDA version causes it or not, but please check that --v2 and --v_parameterization options are appropriate for the model you trained on.

[yetone]

Thank you very much for your prompt reply. My running command is as follows, without specifying the --v2 and --v_parameterization options.

accelerate launch --num_cpu_threads_per_process=2 "train_network.py" --enable_bucket \
--pretrained_model_name_or_path="runwayml/stable-diffusion-v1-5" --train_data_dir="${train_data_dir}" \
--resolution=512,512 --output_dir="${output_dir}" --logging_dir="${logging_dir}" --network_alpha="128" \
--save_model_as=safetensors --network_module=networks.lora --text_encoder_lr=5e-5 --unet_lr=0.0001 \
--network_dim=128 --output_name="${output_name}" --lr_scheduler_num_cycles="1" --learning_rate="0.0001" \
--lr_scheduler="constant" --train_batch_size="2" --max_train_steps="435" --save_every_n_epochs="1" \
--mixed_precision="fp16" --save_precision="fp16" --seed="1234" --caption_extension=".txt" --cache_latents  \
--max_data_loader_n_workers="1" --clip_skip=2 --bucket_reso_steps=64 --xformers \
--use_8bit_adam --bucket_no_upscale

[TingTingin]

Have you tried a different scheduler like cosine with restarts for example?

[yetone]

@TingTingin Thank you for your reply. I don't quite understand the meaning of lr_scheduler, but I will give it a try. Thank you!

[TingTingin]

Try cosine_with_restarts as your scheduler instead

[yetone]

@TingTingin Thanks for your reply! I trained two LoRAs with cosine and cosine_with_restarts lr_scheduler respectively, still splash screen

[TingTingin]

What gpu you have is it a 1660?

[yetone]

@TingTingin My GPU is 3090

[maxcmoi89]

Any update on this issue ? I have the exact same problem.

[TingTingin]

What settings? And gpu

[chouex]

I was using wrong precision by using default accelerate setting (accelerate config default) after setting mixed_precision to fp16 the output lora is working.

!rm ~/.cache/huggingface/accelerate/default_config.yaml
!accelerate config default --mixed_precision fp16