Awesome

Bunny: A family of lightweight multimodal models

📖 Technical report | 🤗 Data | 🤖 Data | 🤗 HFSpace 🐰 Demo

Bunny-Llama-3-8B-V: 🤗 v1.1 | 🤗 v1.0 | 🤗 v1.0-GGUF

Bunny-4B: 🤗 v1.1 | 🤗 v1.0 | 🤗 v1.0-GGUF

Bunny is a family of lightweight but powerful multimodal models. It offers multiple plug-and-play vision encoders, like EVA-CLIP, SigLIP and language backbones, including Llama-3-8B, Phi-3-mini, Phi-1.5, StableLM-2, Qwen1.5, MiniCPM and Phi-2. To compensate for the decrease in model size, we construct more informative training data by curated selection from a broader data source.

We are thrilled to introduce Bunny-Llama-3-8B-V, the pioneering vision-language model based on Llama-3, showcasing exceptional performance. The v1.1 version accepts high-resolution images up to 1152x1152.

Moreover, our Bunny-4B model built upon SigLIP and Phi-3-mini outperforms the state-of-the-art MLLMs, not only in comparison with models of similar size but also against larger MLLMs (7B and 13B). Also, the v1.1 version accepts high-resolution images up to 1152x1152.

News and Updates

• 2024.07.23 🔥 All of the training strategy and data of latest Bunny is released! Check more details about Bunny in Technical Report, Data and Training Tutorial!

• 2024.07.21 🔥 SpatialBot, SpatialQA and SpatialBench are released! SpatialBot is an embodiment model based on Bunny, which comprehends spatial relationships by understanding and using depth information. Try model, dataset and benchmark at GitHub!

• 2024.06.20 🔥 MMR benchmark is released! It is a benchmark for measuring MLLMs' understanding ability and their robustness against misleading questions. Check the performance of Bunny and more details in GitHub!

• 2024.06.01 🔥 Bunny-v1.1-Llama-3-8B-V, supporting 1152x1152 resolution, is released! It is built upon SigLIP and Llama-3-8B-Instruct with S$^2$-Wrapper. Check more details in HuggingFace and wisemodel! 🐰 Demo

• 2024.05.08 Bunny-v1.1-4B, supporting 1152x1152 resolution, is released! It is built upon SigLIP and Phi-3-Mini-4K 3.8B with S$^2$-Wrapper. Check more details in HuggingFace! 🐰 Demo

• 2024.05.01 Bunny-v1.0-4B, a vision-language model based on Phi-3, is released! It is built upon SigLIP and Phi-3-Mini-4K 3.8B. Check more details in HuggingFace! 🤗 GGUF

• 2024.04.21 Bunny-Llama-3-8B-V, the first vision-language model based on Llama-3, is released! It is built upon SigLIP and Llama-3-8B-Instruct. Check more details in HuggingFace, ModelScope, and wisemodel! The GGUF format is in HuggingFace and wisemodel.

• 2024.04.18 Bunny-v1.0-3B-zh, powerful on English and Chinese, is released! It is built upon SigLIP and MiniCPM-2B. Check more details in HuggingFace, ModelScope, and wisemodel! The evaluation results are in the Evaluation. We sincerely thank Zhenwei Shao for his kind help.

• 2024.03.15 Bunny-v1.0-2B-zh, focusing on Chinese, is released! It is built upon SigLIP and Qwen1.5-1.8B. Check more details in HuggingFace, ModelScope, and wisemodel! The evaluation results are in the Evaluation.

• 2024.03.06 Bunny training data is released! Check more details about Bunny-v1.0-data in HuggingFace or ModelScope!

• 2024.02.20 Bunny technical report is ready! Check more details about Bunny here!

• 2024.02.07 Bunny is released! Bunny-v1.0-3B built upon SigLIP and Phi-2 outperforms the state-of-the-art MLLMs, not only in comparison with models of similar size but also against larger MLLMs (7B), and even achieves performance on par with LLaVA-13B! 🤗 Bunny-v1.0-3B

Quickstart

HuggingFace transformers

Here we show a code snippet to show you how to use Bunny-v1.1-Llama-3-8B-V, Bunny-v1.1-4B, Bunny-v1.0-3B and so on with HuggingFace transformers.

This snippet is only used for above models because we manually combine some configuration code into a single file for users' convenience. For example, you can check modeling_bunny_llama.py and configuration_bunny_llama.py and their related parts in the source code of Bunny to see the difference. For other models including models trained by yourself, we recommend loading them with installing the source code of Bunny. Or you can copy files like modeling_bunny_llama.py and configuration_bunny_llama.py into your model and modify auto_map in config.json, but we can't guarantee its correctness and you may need to modify some code to fit your model.

Before running the snippet, you need to install the following dependencies:

pip install torch transformers accelerate pillow

If the CUDA memory is enough, it would be faster to execute this snippet by setting CUDA_VISIBLE_DEVICES=0.

Users especially those in Chinese mainland may want to refer to a HuggingFace mirror site.

import torch
import transformers
from transformers import AutoModelForCausalLM, AutoTokenizer
from PIL import Image
import warnings

# disable some warnings
transformers.logging.set_verbosity_error()
transformers.logging.disable_progress_bar()
warnings.filterwarnings('ignore')

# set device
device = 'cuda'  # or cpu
torch.set_default_device(device)

model_name = 'BAAI/Bunny-v1_1-Llama-3-8B-V' # or 'BAAI/Bunny-Llama-3-8B-V' or 'BAAI/Bunny-v1_1-4B' or 'BAAI/Bunny-v1_0-4B' or 'BAAI/Bunny-v1_0-3B' or 'BAAI/Bunny-v1_0-3B-zh' or 'BAAI/Bunny-v1_0-2B-zh'
offset_bos = 1 # for Bunny-v1_1-Llama-3-8B-V, Bunny-Llama-3-8B-V, Bunny-v1_1-4B, Bunny-v1_0-4B and Bunny-v1_0-3B-zh
# offset_bos = 0 for Bunny-v1_0-3B and Bunny-v1_0-2B-zh

# create model
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16, # float32 for cpu
    device_map='auto',
    trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(
    model_name,
    trust_remote_code=True)

# text prompt
prompt = 'Why is the image funny?'
text = f"A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\n{prompt} ASSISTANT:"
text_chunks = [tokenizer(chunk).input_ids for chunk in text.split('<image>')]
input_ids = torch.tensor(text_chunks[0] + [-200] + text_chunks[1][offset_bos:], dtype=torch.long).unsqueeze(0).to(device)

# image, sample images can be found in https://huggingface.co/BAAI/Bunny-v1_1-Llama-3-8B-V/tree/main/images
image = Image.open('example_2.png')
image_tensor = model.process_images([image], model.config).to(dtype=model.dtype, device=device)

# generate
output_ids = model.generate(
    input_ids,
    images=image_tensor,
    max_new_tokens=100,
    use_cache=True,
    repetition_penalty=1.0 # increase this to avoid chattering
)[0]

print(tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip())

ModelScope

We advise users especially those in Chinese mainland to use ModelScope. snapshot_download can help you solve issues concerning downloading checkpoints.

Before running the snippet, you need to install the following dependencies:

pip install torch modelscope transformers accelerate pillow

If the CUDA memory is enough, it would be faster to execute this snippet by setting CUDA_VISIBLE_DEVICES=0.

import torch
import transformers
from modelscope import AutoTokenizer, AutoModelForCausalLM
from modelscope.hub.snapshot_download import snapshot_download
from PIL import Image
import warnings

# disable some warnings
transformers.logging.set_verbosity_error()
transformers.logging.disable_progress_bar()
warnings.filterwarnings('ignore')

# set device
device = 'cuda'  # or cpu
torch.set_default_device(device)

model_name = 'BAAI/Bunny-Llama-3-8B-V' # or 'BAAI/Bunny-v1.0-3B' or 'BAAI/Bunny-v1.0-3B-zh' or 'BAAI/Bunny-v1.0-2B-zh'
offset_bos = 1 # for Bunny-Llama-3-8B-V and Bunny-v1.0-3B-zh
# offset_bos = 0 for Bunny-v1.0-3B and Bunny-v1.0-2B-zh

# create model
snapshot_download(model_id='thomas/siglip-so400m-patch14-384')
model = AutoModelForCausalLM.from_pretrained(
    model_name,
    torch_dtype=torch.float16, # float32 for cpu
    device_map='auto',
    trust_remote_code=True)
tokenizer = AutoTokenizer.from_pretrained(
    model_name,
    trust_remote_code=True)

# text prompt
prompt = 'Why is the image funny?'
text = f"A chat between a curious user and an artificial intelligence assistant. The assistant gives helpful, detailed, and polite answers to the user's questions. USER: <image>\n{prompt} ASSISTANT:"
text_chunks = [tokenizer(chunk).input_ids for chunk in text.split('<image>')]
input_ids = torch.tensor(text_chunks[0] + [-200] + text_chunks[1][offset_bos:], dtype=torch.long).unsqueeze(0).to(device)

# image, sample images can be found in images folder on https://www.modelscope.cn/models/BAAI/Bunny-Llama-3-8B-V/files
image = Image.open('example_2.png')
image_tensor = model.process_images([image], model.config).to(dtype=model.dtype, device=device)

# generate
output_ids = model.generate(
    input_ids,
    images=image_tensor,
    max_new_tokens=100,
    use_cache=True,
    repetition_penalty=1.0 # increase this to avoid chattering
)[0]

print(tokenizer.decode(output_ids[input_ids.shape[1]:], skip_special_tokens=True).strip())

Model Zoo

Evaluation

The small model with the best performance is denoted as Bunny-v1.0-3B or bunny-phi-2-siglip, whose merged weights can be found here and the LoRA weights can be found here.

We also provide two models that focus on Chinese QA ability, namely Bunny-v1.0-3B-zh (bunny-minicpm-siglip) and Bunny-v1.0-2B-zh (bunny-qwen1.5-1.8b-siglip). The merged weights can be found here and here. The LoRA weights can be found here and here.

Training Tutorial

Install

Either start from our docker or install locally on your own.

Start from Our Docker

Directly start from our configured docker image by docker pull russellrobin/bunny:latest.

1. Run pip install --upgrade transformers && cd Bunny in a running container,

2. Set default GitHub identity by git config user.email "you@example.com" && git config user.name "Your Name",

3. Update Bunny local codes using git pull.

4. pip install -e .

You are all set!

Local Installation

• CUDA and cuDNN

  We use CUDA 11.8 and cuDNN 8.7.0. We actually use the CUDA docker by NVIDIA: docker pull nvcr.io/nvidia/cuda:11.8.0-cudnn8-devel-ubuntu20.04. CUDA 12 is fine, too.

• Create a conda virtual environment and activate it:

  conda create -n bunny python=3.10
  conda activate bunny
  

• Basic requirements

  pip install --upgrade pip  # enable PEP 660 support
  pip install transformers
  pip install torch torchvision xformers --index-url https://download.pytorch.org/whl/cu118
  

• Install apex

  # https://github.com/NVIDIA/apex#from-source
  pip install ninja
  git clone https://github.com/NVIDIA/apex
  cd apex
  # if pip >= 23.1 (ref: https://pip.pypa.io/en/stable/news/#v23-1) which supports multiple `--config-settings` with the same key...
  pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --config-settings "--build-option=--cpp_ext" --config-settings "--build-option=--cuda_ext" ./
  # otherwise
  pip install -v --disable-pip-version-check --no-cache-dir --no-build-isolation --global-option="--cpp_ext" --global-option="--cuda_ext" ./
  

• Install flash-attention

  # https://github.com/Dao-AILab/flash-attention?tab=readme-ov-file#installation-and-features
  pip install packaging
  pip install flash-attn --no-build-isolation
  

• Install bunny and other requirements

  git clone https://github.com/BAAI-DCAI/Bunny.git
  cd Bunny
  pip install -e .
  

Training

Bunny training consists of two stages: (1) pretrain stage: use data to connect a frozen pretrained vision encoder to a frozen LLM, and only the connector is trained; (2) visual instruction tuning stage: use data to teach the model to follow multimodal instructions, where the connector, learnable LLM parameters and vision encoder (optional) are updated.

Bunny is trained on 8 A100 GPUs. Under other circumstances, you can reduce the per_device_train_batch_size and increase the gradient_accumulation_steps accordingly. Always keep the global batch size the same: global_batch_size = per_device_train_batch_size $\times$ gradient_accumulation_steps $\times$ num_gpus.

Support Models

Currently, we support several vision encoders and LLMs.

For vision encoders, we support CLIP, EVA-CLIP and SigLIP.

For LLMs, we support phi-1.5, stablelm-2, qwen1.5-1.8b, minicpm, phi-2, phi-3 and llama3-8b.

Note that there are many variants of above models. We build and test our code based on the exact versions mentioned above. More models will be supported in the future!

Pretrain

• Data preparation

  We use a high-quality coreset with less duplicates and more informative samples of LAION-2B built by this work. We randomly sample 2 million image-text pairs from the coreset and convert them to training format.

  The dataset is available here.

• Run

  Update --model_name_or_path and --vision_tower to the paths of the LLM and vision encoder, respectively. Update MODEL_TYPE and OUTPUT_DIR accordingly. The global batch size is 256. S$^2$-Wrapper would be enabled if --use_s2 True added.

  You may refer to the settings in our experiments in the Training Tutorial.

  sh script/train/pretrain.sh
  

Visual Instruction Tuning

• Data preparation

  We build Bunny-695K by modifying SVIT-mix-665K for finetuning. And we then combine it with LLaVA-665K and ALLaVA-Instruct-4V, i.e., Bunny-LLaVA-1.4M, Bunny-ALLaVA-1.3M, and Bunny-LLaVA-ALLaVA-2M.

  The dataset is available here. If you only want to use Bunny-695K and the related images, you can just download them here.

• Run

  Update --model_name_or_path and --vision_tower to the paths of the LLM and vision encoder, respectively. Update MODEL_TYPE, PRETRAIN_DIR and OUTPUT_DIR accordingly. The global batch size is 128. For MODEL_TYPE = minicpm/phi-3/llama3-8b, change --version to minicpm/phi3/llama, too. S$^2$-Wrapper would be enabled if --use_s2 True added. The vision encoder would be tuned if --unfreeze_vision_tower True added.

  We explore a better strategy including more visual instruction tuning data, S$^2$-Wrapper, trainable vision encoder, weight merging, and etc. You may refer to the settings in our experiments in the Training Tutorial.

  # full-parameter tuning
  sh script/train/finetune_full.sh
  
  # LoRA tuning
  sh script/train/finetune_lora.sh
  

Continuous Fine-tuning

If you want to continuously fine-tuning our released Bunny models on your data or to adapt certain task,

1. Prepare data: convert your data to a JSON file of a list of all samples with the format like Bunny-695K.

2. Prepare model:

   • download Bunny models and if only LoRA provided, merge the LoRA weights and base LLM

     python script/merge_lora_weights.py \
       --model-path /path/to/bunny_lora_weights \
       --model-base /path/to/base_llm_model \
       --model-type phi-2 (or stablelm-2 or phi-1.5 or qwen1.5-1.8b or minicpm or phi-3 or llama3-8b) \
       --save-model-path /path/to/merged_model
     

   • add "continuous_training": true in /path/to/merged_model/config.json to ensure loading the vision tower from merged weights

3. Edit script: both finetune_full.sh and finetune_lora.sh can be used, before:

   • change --model_name_or_path to /path/to/merged_model

   • delete --pretrain_mm_mlp_adapter because we load the cross-modality projector from merged weights

   • customize the hyperparameters, e.g. the learning rate, to fit your dataset

   • for MODEL_TYPE = minicpm/phi-3/llama3-8b, change --version to minicpm/phi3/llama, too. S$^2$-Wrapper would be enabled if --use_s2 True added. The vision encoder would be tuned if --unfreeze_vision_tower True added.

Note that if you continuously fine-tune Bunny models using LoRA, --model-base should be Bunny models rather than the original LLMs when loading.

Demo

Gradio Web UI

• Starting the Controller

  First, start the controller. This service orchestrates communication between the web server and model workers.

  python -m bunny.serve.controller \
  	--host 0.0.0.0 \
  	--port 10000
  

• Launching the Gradio Web Server

  To interact with the models through a web interface, start the Gradio web server.

  Basic start:

  python -m bunny.serve.gradio_web_server \
  	--controller http://localhost:10000 \
  	--model-list-mode reload
  

  If you want to share your web server with others, use --share option. Note that frpc_linux_amd64_v0.2 may be missing and you can fix it following instructions printed on the screen and making the file executable.

  python -m bunny.serve.gradio_web_server \
  	--controller http://localhost:10000 \
  	--model-list-mode reload \
  	--share
  

  Now, you can open the web interface with the URL printed on the screen. You may notice that there is no model in the model list. Do not worry, as we have not launched any model worker yet. It will be automatically updated when you launch a model worker.

• Launching Model Workers

  Model workers handle the processing of model inferences. Configure each worker with the appropriate model and start it. Note to check whether conv_mode is correct here which is decided by the name (path) of the model.

  • For full-parameter tuning models

    python -m bunny.serve.model_worker \
      --host 0.0.0.0 \
      --controller http://localhost:10000 \
      --port 40000 \
      --worker http://localhost:40000 \
      --model-path /path/to/bunny/model \
      --model-type phi-2 (or stablelm-2 or phi-1.5 or qwen1.5-1.8b or minicpm or phi-3 or llama3-8b)
    

  • For LoRA tuning models

    You can use script/merge_lora_weights.py to merge the LoRA weights and base LLM, and use it as above.

    python script/merge_lora_weights.py \
      --model-path /path/to/bunny_lora_weights \
      --model-base /path/to/base_llm_model \
      --model-type phi-2 (or stablelm-2 or phi-1.5 or qwen1.5-1.8b or minicpm or phi-3 or llama3-8b) \
      --save-model-path /path/to/merged_model
    

    Or you can use it without merging as below.

    python -m bunny.serve.model_worker \
      --host 0.0.0.0 \
      --controller http://localhost:10000 \
      --port 40000 \
      --worker http://localhost:40000 \
      --model-path /path/to/bunny_lora_weights \
      --model-base /path/to/base_llm_model \
      --model-type phi-2 (or stablelm-2 or phi-1.5 or qwen1.5-1.8b or minicpm or phi-3 or llama3-8b)
    

CLI Inference (Without Gradio Interface)

For CLI-based inference without using the Gradio interface, use the following command:

• For full-parameter tuning models

  python -m bunny.serve.cli \
  	--model-path /path/to/bunny/model \
  	--model-type phi-2 (or stablelm-2 or phi-1.5 or qwen1.5-1.8b or minicpm) \
  	--image-file /path/to/the/test/image \
  	--conv-mode bunny (change to minicpm/phi3/llama for model-type = minicpm/phi-3/llama3-8b)
  

• For LoRA tuning models

  You can use script/merge_lora_weights.py to merge the LoRA weights and base LLM, and use it as above.

  python script/merge_lora_weights.py \
  	--model-path /path/to/bunny_lora_weights \
  	--model-base /path/to/base_llm_model \
  	--model-type phi-2 (or stablelm-2 or phi-1.5 or qwen1.5-1.8b or minicpm or phi-3 or llama3-8b) \
  	--save-model-path /path/to/merged_model
  

  Or you can use it without merging as below.

  python -m bunny.serve.cli \
  	--model-path /path/to/bunny_lora_weights \
  	--model-base /path/to/base_llm_model \
  	--model-type phi-2 (or stablelm-2 or phi-1.5 or qwen1.5-1.8b or minicpm or phi-3 or llama3-8b) \
  	--image-file /path/to/the/test/image \
  	--conv-mode bunny (change to minicpm/phi3/llama for model-type = minicpm/phi-3/llama3-8b)
  

You can also control temperature, repetition-penalty and max-new-tokens.

Evaluation

For full-parameter tuning models, see evaluation_full.md.

For LoRA tuning models, see evaluation_lora.md.

Citation

If you find this repository helpful, please cite the paper below.

@article{he2024bunny,
      title={Efficient Multimodal Learning from Data-centric Perspective}, 
      author={He, Muyang and Liu, Yexin and Wu, Boya and Yuan, Jianhao and Wang, Yueze and Huang, Tiejun and Zhao, Bo},
      journal={arXiv preprint arXiv:2402.11530},
      year={2024}
}

License

This project utilizes certain datasets and checkpoints that are subject to their respective original licenses. Users must comply with all terms and conditions of these original licenses. The content of this project itself is licensed under the Apache license 2.0.

Acknowledgement

We build our project based on LLaVA: Large Language and Vision Assistant.