Simple implementation of LoRA.
You can use update_model_with_lora on your torch Module and this will automatically update the whole thing. Then you can just save your model as usual and load state dict into original module.
For example:
import torch
from lora import update_model_with_lora
saving_path = "model.pt"
# Create your pretrained model
model = MyComplexNNWithLotsOfParameters(...)
# Update it, such that each linear model will be augmented with LoRA
model = update_model_with_lora(model, rank=5)
# Fine-tune it
train(model)
# Save parameters
torch.save(model.state_dict(), saving_path)
# Now you can use these weights as usual with original model!
model = MyComplexNNWithLotsOfParameters(...)
model.load_state_dict(torch.load(saving_path))