Home

Awesome

LLaMA-Adapter: Efficient Fine-tuning of LLaMA 馃殌

Announcement: We release LLaMA2-Accessory, an open-source toolkit for pre-training, fine-tuning and deployment of LLMs and mutlimodal LLMs.馃敟

Official implementation of 'LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention' and 'LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model'.

<p align="center"> <img src="docs/logo_v4.png"/ width="100%"> <br> </p>

This repo proposes LLaMA-Adapter (V2), a lightweight adaption method for fine-tuning Instruction-following and Multi-modal LLaMA models 馃敟.

Try out the web demo 馃 of LLaMA-Adapter: Hugging Face Spaces, LLaMA-Adapter V2 and ImageBind-LLM.

Join us at Wechat!

News

Released Models

NameApproachDataModalityVisualText
LLaMA-Adapter V1prefix, gateAlpacaText脳LLaMA-7B
LLaMA-Adapter V2 dialogscale, bias, normShareGPTText脳LLaMA-65B
LLaMA-Adapter V2 multimodal[P] prefix, projection, gate <br />[F] bias, norm[P] Image-Text-V1<br />[F] GPT4LLM, LLaVAImage&TextCLIP-ViT-L/14LLaMA-7B
LLaMA-Adapter V2.1 multimodal[P] prefix, projection, gate <br />[F] bias, norm, lora[P] Image-Text-V1<br />[F] GPT4LLM, LLaVA, VQAv2Image&TextCLIP-ViT-L/14LLaMA-7B
ImageBind-LLM[P] prefix, projection, gate<br />[F] bias, norm, lora[P] Image-Text-V1<br />[F] Instruction FollowingImageBind Modalities + Point Cloudimagebind_hugeOpen-Chinese-LLaMA-7B
ImageBind-dialog[P] prefix, projection, gate<br />[F] bias, norm, lora[P] Image-Text-V1<br />[F] LLaVA, ShareGPTImageBind Modalities + Point Cloudimagebind_hugeOpen-Chinese-LLaMA-7B

Overview

Efficiency Comparison:

ModelParametersStorage SpaceTraining Time
Alpaca7B13G3 Hours
LLaMA-Adapter1.2M4.7M1 Hour

By inserting adapters into LLaMA's transformer, our method only introduces 1.2M learnable parameters, and turns a LLaMA into an instruction-following model within 1 hour. For stablizing training at early stages, we propose a novel Zero-init Attention with zero gating mechanism to adaptively incorporate the instructional signals. After fine-tuning, LLaMA-Adapter can generate high-quality instruction-following sentences, comparable to the fully fine-tuned Stanford Alpaca and Alpaca-Lora.

<div align="center"> <img src="docs/pipeline.png"/ width="90%"> </div>

Our approach can be simply extended to Multi-modal Input Instructions. The reasoning framework of image-conditioned LLaMA-Adapter for ScienceQA is as follows, which is also shared by other modalities, such as audio and video.

<div align="center"> <img src="docs/multimodal.png"/ width="90%"> </div>

Setup

Here is a from-scratch script for LLaMA-Adapter V1.

conda create -n llama_adapter -y python=3.8
conda activate llama_adapter

# install pytorch
conda install pytorch cudatoolkit -c pytorch -y

# install dependency and llama-adapter
pip install -r requirements.txt
pip install -e .

Note: To setup other models, please refer to llama_adapter_v2_chat65b, llama_adapter_v2_multimodal7b and imagebind_LLM for more details.

Inference

Please request access to the pre-trained LLaMA from this form (official) or download the LLaMA-7B from Hugging Face (unofficial). Then, obtain the weights of our LLaMA-Adapter from here. We denote the path to the downloaded weights of LLaMA and adapters as TARGET_FOLDER and ADAPTER_PATH.

Here is an example to generate instruction-following sentences with 7B LLaMA model and our LLaMA-Adapter:

torchrun --nproc_per_node 1 example.py \
         --ckpt_dir $TARGET_FOLDER/model_size\
         --tokenizer_path $TARGET_FOLDER/tokenizer.model \
         --adapter_path $ADAPTER_PATH

Training

We release the simple fine-tuning code of LLaMA-Adapter on LLaMA-7B model at here, which is for effortless reproduction with minimal dependencies. We will soon release the fine-tuning code for LLaMA-65B and multi-model LLaMA-Adapter.

Please download the 52K instruction-following training data from Standford Alpaca, and put it under DATA_PATH. Then run:

cd alpaca_finetuning_v1

torchrun --nproc_per_node 8 finetuning.py \
         --model Llama7B_adapter \
         --llama_model_path $TARGET_FOLDER/ \
         --data_path $DATA_PATH/alpaca_data.json \
         --adapter_layer 30 \
         --adapter_len 10 \
         --max_seq_len 512 \
         --batch_size 4 \
         --epochs 5 \
         --warmup_epochs 2 \
         --blr 9e-3 \
         --weight_decay 0.02 \
         --output_dir ./checkpoint/

Comparison with Other Methods

Instruction 1:

How can I develop my critical thinking skills? Give me three suggestions.

  1. Read widely: Reading books, articles, and other sources of information can help you develop your critical thinking skills. By exposing yourself to different perspectives and ideas, you can learn to analyze and evaluate information more effectively.

  2. Practice mindfulness: Mindfulness is the practice of being present in the moment and paying attention to your thoughts and feelings. By being more aware of your thoughts and emotions, you can better understand and evaluate them. This can help you develop a more objective and rational approach to thinking.

  3. Engage in open-minded discussions: Join a discussion group or debate club where you can engage in open-minded discussions with others. This can help you develop your critical thinking skills by challenging your assumptions and encouraging you to consider different perspectives.

  1. Read widely and critically.
  2. Engage in discussions with people who have different perspectives.
  3. Ask questions and seek out answers.

Instruction 2:

As a superhero, how would you explain your origin story to a curious child?

As a superhero, I have a very interesting and exciting origin story to share with you. Once upon a time, there was a young boy named John who lived in a small town. John was always curious and wanted to help others, so he spent his days exploring the world around him and learning about different cultures and people. One day, while he was out exploring, he stumbled upon a mysterious cave. Inside the cave, John found a strange device that had the power to give him incredible abilities. With these abilities, John could fly, shoot lasers from his eyes, and even turn invisible.

John knew that he had to use these powers for good, so he decided to become a superhero and use his abilities to protect his town and the people who lived there. He started to train every day, learning how to use his powers and how to fight against evil. Eventually, John became the superhero known as "The Invisible Man" and he used his powers to help others and make the world a better place.

That's my origin story, and I'm sure you'll love it!

Once upon a time, there was a young boy who had a special power. He used his power to help people in need and protect the innocent. He became known as the superhero, and he used his powers to fight evil and protect the world.

Instruction 3:

Tell me about alpacas.

<details><summary>OTHER METHODS</summary> </details>

Instruction 4:

Tell me about the president of Mexico in 2019.

<details><summary>OTHER METHODS</summary> </details>

Instruction 5:

Write a Python program that prints the first 10 Fibonacci numbers.

def fibonacci():
    a = 0
    b = 1
    for i in range(10):
        a, b = b, a + b
        print(a, b)
    print("Fibonacci numbers: ", a, b)

fibonacci()
<details><summary>OTHER METHODS</summary> 
def fibonacci(n):
    if n == 0:
        return 0
    elif n == 1:
        return 1
    else:
        return fibonacci(n-1) + fibonacci(n-2)

print(fibonacci(10))

numbers = [0, 1]

for i in range(2, 11):
    numbers.append(numbers[i-2] + numbers[i-1])

print(numbers)
# Output: [0, 1, 1, 2, 3, 5, 8, 13, 21, 34]

# Fibonacci numbers are the numbers in the following integer sequence, called the Fibonacci sequence, and characterized by the fact that every number after the first two is the sum of the two preceding ones.

# Fibonacci sequence: 0, 1, 1, 2, 3, 5, 8, 13, 21, 34, ...

# Program:

# Initialize the first two numbers of the Fibonacci sequence
a = 0
b = 1

# Print the first 10 numbers of the Fibonacci sequence
print(a)
print(b)

for i in range(2,10):
    c = a + b
    a = b
    b = c
    print(c)
</details>

Instruction 6:

Write a program that prints the numbers from 1 to 100. But for multiples of three print 'Fizz' instead of the number and for the multiples of five print 'Buzz'. For numbers which are multiples of both three and five print 'FizzBuzz'.

def print_numbers(n):
    for i in range(1, 101):
        if i % 3 == 0 and i % 5 == 0:
            print("FizzBuzz")
        elif i % 3 == 0:
            print("Fizz")
        elif i % 5 == 0:
            print("Buzz")
        else:
            print(i)
<details><summary>OTHER METHODS</summary> 
for i in range(1, 101):
    if i % 3 == 0 and i % 5 == 0:
        print("FizzBuzz")
    elif i % 3 == 0:
        print("Fizz")
    elif i % 5 == 0:
        print("Buzz")
    else:
        print(i)

for i in range(1, 101):
    if i % 3 == 0 and i % 5 == 0:
        print('FizzBuzz')
    elif i % 3 == 0:
        print('Fizz')
    elif i % 5 == 0:
        print('Buzz')
    else:
        print(i)

for (let i = 1; i <= 100; i++) {
  if (i % 3 === 0 && i % 5 === 0) {
    console.log("FizzBuzz");
  } else if (i % 3 === 0) {
    console.log("Fizz");
  } else if (i % 5 === 0) {
    console.log("Buzz");
  } else {
    console.log(i);
  }
}
</details>

Contributors

Peng Gao, Jiaming Han, Renrui Zhang, Aojun Zhou

Hiring Announcement

馃敟 We are hiring interns, postdocs, and full-time researchers at the General Vision Group, Shanghai AI Lab, with a focus on multi-modality and vision foundation models. If you are interested, please contact gaopengcuhk@gmail.com.

Citation

If you find our LLaMA-Adapter code and paper useful, please kindly cite:

@article{zhang2023llamaadapter,
  title = {LLaMA-Adapter: Efficient Fine-tuning of Language Models with Zero-init Attention},
  author={Zhang, Renrui and Han, Jiaming and Liu, Chris and Gao, Peng and Zhou, Aojun and Hu, Xiangfei and Yan, Shilin and Lu, Pan and Li, Hongsheng and Qiao, Yu},
  journal={arXiv preprint arXiv:2303.16199},
  year={2023}
}

If you find our LLaMA-Adapter V2 code and paper useful, please kindly cite:

@article{gao2023llamaadapterv2,
  title = {LLaMA-Adapter V2: Parameter-Efficient Visual Instruction Model},
  author={Gao, Peng and Han, Jiaming and Zhang, Renrui and Lin, Ziyi and Geng, Shijie and Zhou, Aojun and Zhang, Wei and Lu, Pan and He, Conghui and Yue, Xiangyu and Li, Hongsheng and Qiao, Yu},
  journal={arXiv preprint arXiv:2304.15010},
  year={2023}
}

Acknowledgement

This repo benefits from LLaMA, Stanford Alpaca, and Alpaca-Lora. Thanks for their wonderful works.