modules_to_save Incorrect Overlap in Multiple LoRA Adapters

[saeid93]

System Info

Python 3.11.9
transformers==4.40.2
peft==0.11.2

Who can help?

@BenjaminBossan
A bug occurs in the PEFT library when using multiple LoRA adapters, each with a unique modules_to_save configuration. The issue arises when the modules_to_save from the first LoRA adapter (e.g., adapter_1) is applied to subsequent adapters (e.g., adapter_2), rather than maintaining independent configurations. As a result, modules specified in modules_to_save for adapter_1 also appear in adapter_2, leading to unintended behavior and possibly affecting fine-tuning accuracy. This incorrect handling of modules_to_save causes duplicate entries where only the respective LoRA adapter’s modules should be saved.

Information

• The official example scripts
• My own modified scripts

Tasks

• An officially supported task in the examples folder
• My own task or dataset (give details below)

Reproduction

The following example code demonstrates this issue, displaying the model structure where adapter_2 contains modules meant only for adapter_1.

Example Code

import os
from transformers import AutoModelForCausalLM
from peft import LoraConfig, get_peft_model, PeftModel

# Get the directory of the current Python script
script_dir = os.path.dirname(os.path.abspath(__file__))

# Define relative paths for adapters
adapter_1_path = os.path.join(script_dir, "adapter_1")
adapter_2_path = os.path.join(script_dir, "adapter_2")

# Load base model
base_model = AutoModelForCausalLM.from_pretrained("gpt2")

# Define LoRA configs with different modules_to_save
lora_config_1 = LoraConfig(
    r=8,
    lora_alpha=32,
    target_modules=["c_attn"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    modules_to_save=["lm_head"]
)

lora_config_2 = LoraConfig(
    r=8,
    lora_alpha=32,
    target_modules=["c_proj"],
    lora_dropout=0.05,
    bias="none",
    task_type="CAUSAL_LM",
    modules_to_save=["wte"]
)

# Apply and save the first adapter
os.makedirs(adapter_1_path, exist_ok=True)
model_with_lora_1 = get_peft_model(base_model, lora_config_1, adapter_name="adapter_1")
model_with_lora_1.save_pretrained(adapter_1_path)

# Apply and save the second adapter
os.makedirs(adapter_2_path, exist_ok=True)
model_with_lora_2 = get_peft_model(base_model, lora_config_2, adapter_name="adapter_2")
model_with_lora_2.save_pretrained(adapter_2_path)

# Load a fresh base model and wrap it in PeftModel by loading the first adapter
base_model = AutoModelForCausalLM.from_pretrained("gpt2")
peft_model = PeftModel.from_pretrained(base_model, os.path.join(adapter_1_path, "adapter_1"), adapter_name="adapter_1")

# Load the second adapter into the PeftModel
peft_model.load_adapter(os.path.join(adapter_2_path, "adapter_2"), adapter_name="adapter_2")

# Display structure and inspect unexpected 'modules_to_save' overlap
print("Expected `modules_to_save` for each adapter:")
print("Adapter 1 `modules_to_save`: ['lm_head']")
print("Adapter 2 `modules_to_save`: ['wte']")
print("\nActual model structure and `modules_to_save` contents:\n")
print(peft_model.transformer.wte)
print(peft_model.lm_head)

The code output will be:

Expected `modules_to_save` for each adapter:
Adapter 1 `modules_to_save`: ['lm_head']
Adapter 2 `modules_to_save`: ['wte']

Actual model structure and `modules_to_save` contents:

ModulesToSaveWrapper(
  (original_module): Embedding(50257, 768)
  (modules_to_save): ModuleDict(
    (adapter_2): Embedding(50257, 768)
  )
)
ModulesToSaveWrapper(
  (original_module): Linear(in_features=768, out_features=50257, bias=False)
  (modules_to_save): ModuleDict(
    (adapter_1): Linear(in_features=768, out_features=50257, bias=False)
    (adapter_2): Linear(in_features=768, out_features=50257, bias=False)
  )
)

Expected behavior

As you see adapter 2 is also built for the "lm_head" module to which it shouldn't, the expected output is shown below:

Expected `modules_to_save` for each adapter:
Adapter 1 `modules_to_save`: ['lm_head']
Adapter 2 `modules_to_save`: ['wte']

Actual model structure and `modules_to_save` contents:

ModulesToSaveWrapper(
  (original_module): Embedding(50257, 768)
  (modules_to_save): ModuleDict(
    (adapter_2): Embedding(50257, 768)
  )
)
ModulesToSaveWrapper(
  (original_module): Linear(in_features=768, out_features=50257, bias=False)
  (modules_to_save): ModuleDict(
    (adapter_1): Linear(in_features=768, out_features=50257, bias=False)
  )
)

[BenjaminBossan]

Thanks a lot for reporting this. Indeed, the handling of modules_to_save can be messy at times and the outcome you show should be avoided. I don't have the opportunity to test this right now, but my assumption is that this extra module won't disrupt the results for adapter 2 because it is a copy of the original layer and behaves exactly the same, as that right?

[saeid93]

No worries, glad to be of any help. As far as I have tested it should be fine and using the correct loaded layer, the only problem is redundancy in loaded modules. I also dug a bit deeper and noticed that the problem originates from this function:

peft/src/peft/peft_model.py

Line 966 in 162d7e5

def set_additional_trainable_modules(self, peft_config, adapter_name):

For an unknown reason when using load_adapter:

peft/src/peft/peft_model.py

Line 969 in 162d7e5

self.modules_to_save = set(peft_config.modules_to_save)

The set is not being updated to only the new layer and it will still hold the old layer in the set too (which shouldn't). For example if I manually hack the above script the problem will be solved:

...

# Apply and save the second adapter
os.makedirs(adapter_2_path, exist_ok=True)
model_with_lora_2 = get_peft_model(base_model, lora_config_2, adapter_name="adapter_2")
model_with_lora_2.save_pretrained(adapter_2_path)

# Load a fresh base model and wrap it in PeftModel by loading the first adapter
base_model = AutoModelForCausalLM.from_pretrained("gpt2")
peft_model = PeftModel.from_pretrained(base_model, os.path.join(adapter_1_path, "adapter_1"), adapter_name="adapter_1")

peft_model.modules_to_save = {"wte"} # <----------- HERE manually changing the modules_to_save
# Load the second adapter into the PeftModel
peft_model.load_adapter(os.path.join(adapter_2_path, "adapter_2"), adapter_name="adapter_2")

...

[BenjaminBossan]

Okay, I managed to reproduce the error. My tentative fix is in #2220. Right now, there is a CI issue but it should hopefully resolve itself soon. Meanwhile, it would be great if you could check if the fix makes sense to you.

Just a side note, when adding multiple adapters, don't use get_peft_model twice, use peft_model.add_adapter instead. But even with that change, the problem is reproducible.