OLA-VLM

Jitesh Jain^*, Zhengyuan Yang, Humphrey Shi^†, Jianfeng Gao^†, Jianwei Yang^†

^*Work done during an internship at Microsoft Research, Redmond    ^†Equal Advising

[Project Page] | [arXiv] [Model Checkpoints] [Video] [BibTeX]

This repo contains the code for our paper OLA-VLM: Elevating Visual Perception in Multimodal LLMs with Auxiliary Embedding Distillation.

We propose distilling target visual information into the intermediate representations of the LLM from a set of target encoders. We adopt a predictive embedding optimization approach at selected LLM layers during training to minimize the embedding losses along with the next token prediction (NTP) objective, resulting in a vision-centric approach to training the Multimodal Large Language Model.

Contents

1. Installation Instructions
2. Demo
3. Getting Started
4. Results
5. Citation

News

• [December 12, 2024]:
  • Project Page, ArXiv Preprint and GitHub Repo are public! 🚀
  • We also open-source the model checkpoints and probes on huggingface hub! 🎁

Installation Instructions

Note: We trained all our models on AMD MI300x GPUs. However, in this repo, we provide instructions for Nvidia GPUs considering their wider usage.

• Clone this repository.

  git lfs install
  git clone https://github.com/SHI-Labs/OLA-VLM
  cd OLA-VLM

• Setup conda environment with the base dependencies.

  conda create -n ola_vlm -y
  conda activate ola_vlm
  pip install -e .
  pip install flash-attn --no-build-isolation
  pip install scikit-learn icecream datasets pytorch-fid lpips opencv-python-headless
  pip install setuptools==61.0.0
  pip install -e lmms-eval/
  pip install huggingface_hub==0.24.7
  pip install transformers==4.41.1

Demo

You can use the Gradio interface to interact with OLA-VLM locally. The demo also supports visualizing the respresentations from the slected intermediate LLM layers (embedding loss positions).

# install demo-specific libraries
pip install -e .["demo"]

# start the demo
CUDA_VISIBLE_DEVICES=0 python demo.py --model-path shi-labs/pretrain_dsg_OLA-VLM-CLIP-ViT-Llama3-8b --PT-model-path shi-labs/pretrain_dsg_OLA-VLM-CLIP-ViT-Llama3-8b

Getting Started

Note: We provide the guide to integrating the embeddding losses from OLA-VLM into any custom MLLM in Custom_MLLM.md

Training

• Please see Training.md for training commands and dataset preparation.
• We train all our models using 16 192G MI300X AMD GPUs.

Evaluation

Please see Evaluation.md for evaluation commands

Probing

Please see Probing.md for probing commands.

Results

Method	Training Stages	LLM	Base Encoder	CV-Bench	MMStar	RWQA	OK-VQA	Checkpoint
OLA-VLM	PT + IFT	Phi3-4k-mini	CLIP-ViT-L	62.5	36.0	58.0	56.4	ckpt
OLA-VLM	PT + IFT	Phi3-4k-mini	CLIP-ConvNeXT-XXL	63.9	38.4	58.4	56.5	ckpt
OLA-VLM	PT + IFT	Llama3-8b	CLIP-ViT-L	61.4	39.5	57.9	56.6	ckpt
OLA-VLM	PT + IFT	Llama3-8b	CLIP-ConvNeXT-XXL	61.5	38.5	55.0	59.0	ckpt
OLA-VLM	PT + VPT + IFT	Llama3-8b	CLIP-ConvNeXT-XXL	64.6	40.6	62.9	61.1	ckpt

Citation

If you found OLA-VLM useful in your research, please consider starring ⭐ us on GitHub and citing 📚 us in your research!

@article{jain2024ola_vlm,
      title={{OLA-VLM: Elevating Visual Perception in Multimodal LLMs with Auxiliary Embedding Distillation}},
      author={Jitesh Jain and Zhengyuan Yang and Humphrey Shi and Jianfeng Gao and Jianwei Yang},
      journal={arXiv},
      year={2024}
}

Acknowledgement

We thank the authors of LLaVA-1.5, OneFormer, Depth-Anything v2, and unCLIP-SD for open-sourcing their codebase and checkpoints. We are grateful to the authors of cambrian and MMStar for releasing their code for CV-Bench and MMStar evaluation, respectively.