This is not directly possible. What you could try is to add all the layers you want to eventually train into target_modules. Then, you go through the modules and manually disable the gradient of those you don't want to train:

target_modules = [
    "0.attention.self.query", "0.attention.self.value",
    "11.attention.self.query", "11.attention.self.value",
]
...
model = get_peft_model(...)
for name, module in model.named_modules():
    if name ...:
        module.requires_grad_(False)

I've never tried this, so not 100% sure if it'll work, but worth a try.

Actually, that's what I am using. But model (just fine-tune the last transformer) created by this way performs differently with model created directly (pass the target_modules argument). I guess it's because of the initialization of additional layers changes the random number sequence.

Yes, if you create more LoRA layers, the random seed will be ticked more often. But I don't understand what the big issue would be with that. Maybe you could show your code and explain what you expect vs what actually happens.

Just to know whether it's the reason of random seed.

Thanks a lot for your reply.