Fix: Multiple adapters with bnb layers

[BenjaminBossan]

Resolves #1239

Description

Fixes a bug that led to an error when loading multiple adapters into a
peft model that uses bnb layers.

Notes

I tested this locally on GPU and tests pass for me.

Implementation

While working on this, I thought we could make a simplification to our
layer code. Right now, we have code like this inside of
_create_and_replace:

    if isinstance(target, conv2d):
        target.update_layer_conv2d(...)
    elif isinstance(target, embedding):
        target.update_layer_embedding(...)
    elif ...

If we move the update_layer* methods from LoraLayer to the subclasses,
i.e. Linear, Embedding, etc., and give them all the same name (i.e.
update_layer_embedding => update_layer), we could simplify the code
above to:

    if isinstance(target, LoraLayer):
        target.update_layer(...)

Additionally, it makes more sense to me that the code for
update_layer_embedding lives inside of Embedding, not the parent class.
If you agree, I could work on that in a separate PR.

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[younesbelkada]

Thank you !

[chris111110]

while I tried to use this solution to load multiple lora in peft model, it appear a error.

245 parent, target, target_name = _get_submodules(model, key)
247 optional_kwargs = {
248 "loaded_in_8bit": getattr(model, "is_loaded_in_8bit", False),
249 "loaded_in_4bit": getattr(model, "is_loaded_in_4bit", False),
250 "current_key": key,
...
1672 buffers[name] = value
1673 else:
-> 1674 super().setattr(name, value)

AttributeError: can't set attribute 'weight'

and it seems to work if I modify the code in _create_new_module this way:

elif AutoGPTQQuantLinear is not None and isinstance(target_base_layer, AutoGPTQQuantLinear):
new_module = QuantLinear(target, adapter_name, **kwargs)
###xlc
if target == None:
target.weight = target_base_layer.qweight
else:
target = target_base_layer
target.weight = target_base_layer.qweight

[BenjaminBossan]

@ChrisXULC Thanks for reporting. Do you have some code to reproduce the error?

[pacman100]

Thank you @BenjaminBossan for fixing the issue when using multiple adapters with bnb layers, nice tests! LGTM! 🚀

[chris111110]

@ChrisXULC Thanks for reporting. Do you have some code to reproduce the error?

@BenjaminBossan it could be reproduce while trying to load multiple loras into Qwen 7b int 4 model. before loading loras, I have modified the anaconda3/envs/lib/python3.10/site-packages/peft/tuners/lora/model.py like you did.

model = AutoModelForCausalLM.from_pretrained(
"/home/models/qwen/Qwen-7B-Chat-Int4",
device_map="auto",
trust_remote_code=True
# use_cache_quantization=True,
# use_cache_kernel=True,
# use_flash_attn=False
).eval()

model = PeftModel.from_pretrained(model,'/home//models/quantized_model_lora', adapter_name="23",device_map="cuda", load_in_4bit=True)
model.load_adapter('/home/lora_all/quantized_model_vul',adapter_name = '22')

[BenjaminBossan]

@ChrisXULC Thanks a lot. The issue should be fixed with the latest commit. I omitted the if target == None: guard in from your suggestion, I don't think that can happen, but LMK if I'm wrong.

(Note: tests are failing because of #1252)

[chris111110]

@BenjaminBossan I have changed it to target.qweight=target_base_layer.qweight，I think it should be ok

[BenjaminBossan]

@ChrisXULC Thanks for checking. If you also have time to check PEFT from this branch, that would be fantastic.

[chris111110]

@BenjaminBossan cool!

[BenjaminBossan]

Tests are green now after merging #1252, I'll merge this one too.