about training

[megake98]

I’d like to thank you for open-sourcing this project， the image restoration results are truly impressive.

And how should the training script be configured to prevent OOM errors? Is the original training code designed for Accelerate DDP training? I switched it to Zero-2, but I am still encountering OOM errors.

[double8fun]

May I ask what your GPU specs are?

[megake98]

May I ask what your GPU specs are?
H800, 80G

[double8fun]

80G should be just enough for our default training configuration.
If you still encounter OOM issues, you might consider preprocessing the prompt embedding for the discriminator Wan to eliminate the need for the text encoder, or appropriately reducing the lora rank of the generator.