[AdaLora] Fix adalora inference issue

[younesbelkada]

Fixes #571

As stated by @BenjaminBossan the regularization term is only computed in training mode, i.e. when outputs.loss is different than None

Can confirm this snippet that is not working on main

import torch
from peft import AdaLoraConfig, get_peft_model, AutoPeftModelForCausalLM
from transformers import AutoModelForCausalLM 

model_id = "facebook/opt-350m"
peft_config = AdaLoraConfig(
    task_type="CAUSAL_LM"
)

model = AutoModelForCausalLM.from_pretrained(model_id)
dummy_input = torch.LongTensor([[1, 2, 3, 4, 5]])

model = get_peft_model(model, peft_config)
model.save_pretrained("test-adalora")

model = AutoPeftModelForCausalLM.from_pretrained("test-adalora")
_ = model(dummy_input)

Now works

Also fixes an issue when creating an adalora model with dropout == 0.

[HuggingFaceDocBuilderDev]

The documentation is not available anymore as the PR was closed or merged.

[BenjaminBossan]

LGTM, thanks!

The trainable_adapter_name parameter is also used in this method:

peft/src/peft/tuners/adalora.py

Line 297 in b15c185

def update_and_allocate(self, global_step):

But it probably makes no sense to call it in inference mode, so I guess we can leave it as is.