Original model description:

license: mit language: - en pipeline_tag: text-generation tags: - 'quantization ' - lora

LoftQ Initialization

LoftQ (LoRA-fine-tuning-aware Quantization) provides a quantized backbone Q and LoRA adapters A and B, given a full-precision pre-trained weight W.

This model, phi-2-4bit-64rank, is obtained from phi-2. The backbone is under LoftQ/phi-2-4bit-64rank and LoRA adapters are under the subfolder='loftq_init'.

Model Info

Backbone

Stored format: torch.float16
Size: ~ 5.5 GiB
Loaded format: bitsandbytes nf4
Size loaded on GPU: ~1.4 GiB

LoRA adapters

rank: 64
lora_alpha: 16
target_modules: ["q_proj", "k_proj", "v_proj", "dense", "fc1", "fc2"]

Usage

Training Here's an example of loading this model and preparing for the LoRA fine-tuning.

import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

MODEL_ID = "LoftQ/phi-2-4bit-64rank"

base_model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, 
    torch_dtype=torch.float32,  # you may change it with different models
    quantization_config=BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.float32,  # float32 is tested and veryfied
        bnb_4bit_use_double_quant=False,
        bnb_4bit_quant_type='nf4',
    ),
)
peft_model = PeftModel.from_pretrained(
    base_model,
    MODEL_ID,
    subfolder="loftq_init",
    is_trainable=True,
)

# Do training with peft_model ...

Experiment Results

We have conducted experiments on supervised fine-tuning of GSM8K.

Model	Bits	Rank	LoRA Initial	GSM8K
Phi-2	16	-	Full model fine-tuning	66.8±1.2
Phi-2	16	64	Gaussian + 0 (LoRA)	64.8±0.5
Phi-2	4	64	Gaussian + 0 (QLoRA)	60.2±0.6
Phi-2	4	64	LoftQ	64.1±0.7

Inference Here is an example code for inference after the model has been fine-tuned on GSM8K.

import torch
from transformers import AutoModelForCausalLM, BitsAndBytesConfig
from peft import PeftModel

MODEL_ID = "LoftQ/phi-2-4bit-64rank"

base_model = AutoModelForCausalLM.from_pretrained(
    MODEL_ID, 
    torch_dtype=torch.float32,  # you may change it with different models
    quantization_config=BitsAndBytesConfig(
        load_in_4bit=True,
        bnb_4bit_compute_dtype=torch.float32,  # float32 is tested and veryfied
        bnb_4bit_use_double_quant=False,
        bnb_4bit_quant_type='nf4',
    ),
)
peft_model = PeftModel.from_pretrained(
    base_model,
    MODEL_ID,
    subfolder="gsm8k",
    is_trainable=True,
)

# Do inference with peft_model ...

See the full code at our Github Repo

Citation

@article{li2023loftq,
  title={Loftq: Lora-fine-tuning-aware quantization for large language models},
  author={Li, Yixiao and Yu, Yifan and Liang, Chen and He, Pengcheng and Karampatziakis, Nikos and Chen, Weizhu and Zhao, Tuo},
  journal={arXiv preprint arXiv:2310.08659},
  year={2023}
}