LLM-Adapters

LLM-Adapters: An Adapter Family for Parameter-Efficient Fine-Tuning of Large Language Models

LLM-Adapters is an easy-to-use framework that integrates various adapters into LLMs and can execute adapter-based PEFT methods of LLMs for different tasks. LLM-Adapter is an extension of HuggingFace's PEFT library, many thanks for their amazing work! Please find our paper at this link: https://arxiv.org/abs/2304.01933.

The framework includes state-of-the-art open-access LLMs: LLaMa, OPT, BLOOM, and GPT-J, as well as widely used adapters such as Bottleneck adapters, Parallel adapters, and LoRA.

Supported Adapters:

LoRA: LORA: LOW-RANK ADAPTATION OF LARGE LANGUAGE MODELS
AdapterH: Parameter-Efficient Transfer Learning for NLP
AdapterP: GMAD-X: An Adapter-Based Framework for Multi-Task Cross-Lingual Transfer
Parallel: TOWARDS A UNIFIED VIEW OF PARAMETER-EFFICIENT TRANSFER LEARNING
Prefix Tuning: Prefix-Tuning: Optimizing Continuous Prompts for Generation, P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks
P-Tuning: GPT Understands, Too
Prompt Tuning: The Power of Scale for Parameter-Efficient Prompt Tuning

Latest News 🔥🔥

[2023-04-21] We released math10k dataset and the LLaMA-13B adapter checkpoints. The LLaMA-13B-Parallel model achieves 91% of GPT-3.5 performance!
[2023-04-10] We can support GPT-Neo and ChatGLM now!
[2023-04-04] Release code and dataset

Setup

Install dependencies

pip install -r requirements.txt
cd peft/
pip install -e .

Set environment variables, or modify the files referencing BASE_MODEL:

# Files referencing `BASE_MODEL`
# export_hf_checkpoint.py
# export_state_dict_checkpoint.py

export BASE_MODEL=yahma/llama-7b-hf

Both finetune.py and generate.py use --base_model flag as shown further below.

If bitsandbytes doesn't work, install it from source. Windows users can follow these instructions.

Training(finetune.py)

This file contains some code related to prompt construction and tokenization.In this file, specify different adapters and different sets of data, so that different models can be trained.

Example usage for multiple GPUs:

WORLD_SIZE=2 CUDA_VISIBLE_DEVICES=0,1 torchrun --nproc_per_node=2 --master_port=3192 finetune.py \
  --base_model 'yahma/llama-7b-hf' \
  --data_path 'math_data.json' \
  --output_dir './trained_models/llama-lora' \
  --batch_size 16 \
  --micro_batch_size 4 \
  --num_epochs 3 \
  --learning_rate 3e-4 \
  --cutoff_len 256 \
  --val_set_size 120 \
  --adapter_name lora

The math_data.json file contains preprocessed instruction data from the addsub, SingleEQ, MultiArith, AQuA, SVAMP and GSM8K dataset. yahma/llama-7b-hf is a base model, LLaMa-7B. Add lora adapter to this model.

Example usage for Single GPUs:

CUDA_VISIBLE_DEVICES=0 python finetune.py \
  --base_model 'yahma/llama-7b-hf' \
  --data_path 'math_data.json' \
  --output_dir './trained_models/llama-lora' \
  --batch_size 16 \
  --micro_batch_size 4 \
  --num_epochs 3 \
  --learning_rate 3e-4 \
  --cutoff_len 256 \
  --val_set_size 120 \
  --adapter_name lora

Moreover, you can use --use_gradient_checkpointing to save more GPU memory, but it will increase the training time.

To use the AdapterH, just add the following arguments:

--adapter_name bottleneck # use the bottleneck adapter, refers to AdapterH in the result table

To use the AdapterP, just add the following arguments:

--adapter_name bottleneck 
--use_adapterp  # use the AdapterP, refers to AdapterP in the result table

To use parallel adapter, just add the following arguments:

--adapter_name bottleneck
--use_parallel_adapter

Note that, In order to facilitate INT8 training of large models with parallel adapters, we have adopted a technique whereby the parallel adapter layers are incorporated into multi-head attention layers and MLP layers, in parallel with Linear layers. It is different from Hu et al. (2021).

Inference (generate.py)

This file reads the foundation model from the Hugging Face model hub and the LoRA weights from './trained_models/llama-lora' , and runs a Gradio interface for inference on a specified input. Users should treat this as example code for the use of the model, and modify it as needed. Example usage:

CUDA_VISIBLE_DEVICES=0 torchrun generate.py \
    --base_model 'yahma/llama-7b-hf' \
    --lora_weights './trained_models/llama-lora'

Evaluation (evaluate.py)

To evaluate the performance of the finetuned model on the Arithmetic Reasoning tasks, you can use the following command:

CUDA_VISIBLE_DEVICES=0 python evaluate.py 
    --model LLaMA-7B \ #specify the base model
    --adapter LoRA \   #specify the adapter name ["LoRA", "AdapterH", "AdapterP", "Parallel"， "Scaled_Parallel""]
    --dataset SVAMP \  #specify the test dataset
    --base_model 'yahma/llama-7b-hf' \
    --lora_weights './trained_models/llama-lora'

Resource Consumption

There is a table of resouce needed for different adapters, which contains Trainable Parameters, GPU RAM Usage, and Fine-tuning Time on the Arithmetic Reasoning dataset math_data.json

Hyper-parameter setting: num_epochs=3, lora_r=8, lora_alpha=16, bottleneck_size=256

Models: LLaMA-13B, LLaMA-7B, BLOOM-6.7B, GPT-j-6B Dataset: 3.2K math word problems

Hardware: 2*3090 GPUs

Model	Trainable Parameters	GPU RAM Usage	Fine-tuning Time
LLaMA-7B-LoRA	4.2M	18GB	1h
LLaMA-7B-AdapterH	200M	22GB	1h
LLaMA-7B-AdapterP	200M	22GB	1h
LLaMA-7B-Parallel	200M	22GB	1h

Finetune Result

There is a finetune result in different model with six dataset, which contains MultiArith, GSM8K, AddSub, AQuA, SingleEq, SVAMP

Model	Params	MultiArith	GSM8K	AddSub	AQuA	SingleEq	SVAMP	Average
GPT-3.5	-	83.8	56.4	85.3	38.9	88.1	69.9	70.4
LLaMA-13B-LoRA	6.5M	93.3	43.3	80.0	20.5	84.6	52.9	62.4
LLaMA-13B-AdapterH	314M	94.0	36.1	82.3	19.7	84.8	52.9	61.6
LLaMA-13B-AdapterP	104M	94.8	41.0	81.3	19.3	87.0	51.1	62.4
LLaMA-13B-Parallel	314M	95.0	43.8	84.6	20.9	88.0	53.5	64.3
LLaMA-7B-LoRA	4.2M	88.3	30.9	78.5	14.2	74.8	47.2	55.7
LLaMA-7B-AdapterH	200M	93.8	29.8	70.6	16.1	71.1	37.7	53.2
LLaMA-7B-AdapterP	66M	91.0	30.2	75.7	14.9	75.4	43.3	55.1
LLaMA-7B-Parallel	200M	93.7	33.3	80.5	16.5	81.7	46.5	58.7
BLOOM-7B-LoRA	4M	73.0	9.9	41.8	16.9	40.7	25.1	34.6
BLOOM-7B-AdapterH	125M	81.8	16.5	76.5	18.9	71.3	37.8	50.5
BLOOM-7B-AdapterP	62M	87.7	18.0	69.6	20.9	68.3	32.1	49.4
BLOOM-7B-Parallel	125M	78.2	15.7	65.4	20.5	64.2	35.1	46.5
GPT-j-6B-LoRA	3.7M	80.5	17.4	74.9	18.1	72.2	43.8	51.2
GPT-j-6B-AdapterH	117M	82.5	17.9	83.8	21.3	76.8	40.0	53.7
GPT-j-6B-AdapterP	58M	90.3	19.1	80.7	18.5	81.3	41.3	55.2
GPT-j-6B-Parallel	176M	77.8	17.5	77.2	20.5	74.8	39.8	51.3

Adapter support matrix

This metrix shows whether different models can use LoRA,AdapterH,AdapterP,Parallel and Scaled Parallel adapters.

Adapter	LoRA	AdapterH	AdapterP	Parallel	Prefix Tuning	P-Tuning	Prompt Tuning
LLaMA	✅	✅	✅	✅	✅	✅	✅
BLOOM	✅	✅	✅	✅	✅	✅	✅
GPT-J	✅	✅	✅	✅	✅	✅	✅
OPT	✅	✅	✅	✅	✅	✅	✅
GPT-2	✅	🔧Developing	🔧Developing	🔧Developing	✅	✅	✅
GPT-Neo	✅	✅	✅	✅	✅	✅	✅
GPT-NeoX-20B	✅	🔧Developing	🔧Developing	🔧Developing	✅	✅	✅
ChatGLM	✅	✅	✅	✅	✅	✅	✅

TODO List

⭐ Star History

Citing LLM-Adapter

If you use LLM-Adapters in your publication, please cite it by using the following BibTeX entry.

@article{hu2023llm,
  title={LLM-Adapters: An Adapter Family for Parameter-Efficient Fine-Tuning of Large Language Models},
  author={Hu, Zhiqiang and Lan, Yihuai and Wang, Lei and Xu, Wanyu and Lim, Ee-Peng and Lee, Roy Ka-Wei and Bing, Lidong and Poria, Soujanya},
  journal={arXiv preprint arXiv:2304.01933},
  year={2023}
}

Acknowledgement

This repo benefits from PEFT, Adapter-Transformer, Alpaca-lora. Thanks for their wonderful works. Additionally, we thank DONG Shan and dream.ai for the exceptional logo design, which has added immense value to our project.

marcovirgolin/LLM-Adapters