📃 Paper • 🌐 Demo • 🤗 LongLLaVA-53B-A13B • 🤗 LongLLaVA-9B

🌈 Update

• [2024.09.05] LongLLaVA repo is published！🎉
• [2024.10.12] LongLLaVA-53B-A13B, LongLLaVA-9b and Jamba-9B-Instruct are repleased！🎉

Architecture

Click to view the architecture image

Results

Click to view the Results

• Main Results
• Diagnostic Results
• Video-NIAH

Results reproduction

1. Environment Setup

pip install -r requirements.txt

2. Data DownLoad and Construction

Dataset Taxonomy

• Dataset DownLoading and Construction

  Coming Soon.

3. Training

• Downloading Language Models

  🤗 Jamba-9B-Instruct

• Stage I: Single-image Alignment.

  bash Align.sh

• Stage II: Single-image Instruction-tuning.

  bash SingleImageSFT.sh

• Stage III: Multi-image Instruction-tuning.

  bash MultiImageSFT.sh

4. Evaluation

• Command Line Interface

python cli.py --model_dir path-to-longllava

• Model Inference

query = 'What does the picture show?'
image_paths = ['image_path1'] # image or video path

from cli import Chatbot
bot = Chatbot(path-to-longllava)
output = bot.chat(query, image_paths)
print(output) # Prints the output of the model

• Benchmarks

python Eval.sh

5. Reproduce other results in Paper

• FLOPs

python /utils/cal_flops.py

• Prefill Time & Throughput & GPU Memory Usage

python ./benchmarks/Efficiency/evaluate.py
python ./benchmarks/Efficiency/evaluatevllm.py

• DownCycling To Transfer Jamba-MoE to Dense

python ./utils/dense_downcycling.py

TO DO

• Release Data Construction Code

Acknowledgement

• LLaVA: Visual Instruction Tuning (LLaVA) built towards GPT-4V level capabilities and beyond.

Citation

@misc{wang2024longllavascalingmultimodalllms,
      title={LongLLaVA: Scaling Multi-modal LLMs to 1000 Images Efficiently via Hybrid Architecture}, 
      author={Xidong Wang and Dingjie Song and Shunian Chen and Chen Zhang and Benyou Wang},
      year={2024},
      eprint={2409.02889},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2409.02889}, 
}