zhuoyanxu/ada-llava-v1.5-13b

Ada-LLaVA Model Card

Ada-LLaVA 13B is an open-source adaptive inference framework for multimodal Large Language Models (MLLMs) that dynamically adjusts its operations based on available computational resources and latency requirements.

See the paper for more details: Learning to Inference Adaptively for Multimodal Large Language Models

Model details: https://zhuoyan-xu.github.io/ada-llava/

Model Details

Model Type: Ada LLaVA 13B follows the LLaVA-v1.5 stage-2 training pipeline, with CLIP-ViT-L-336px as visual encoder (336*336 image resolution), Vicuna-v1.5-13B as base LLM and a two-layer MLP as vision-language connector, customized embedding model and MLP as latency scheduler.

It was trained with stage-2 pipeline as LLaVA:

Instruction tuning: Freeze vision encoder, train the remaining model with multimodal instruction following data of tabular and non-tabular tasks.

Code Base: We use the official code of LLaVA-v1.5 for model training and inference, and the saved model checkpoint is uploaded to this repository.

Model Date: Ada-LLaVA 13B was trained in Oct 2024.

License

AdaLLaVA is based on LLaVA-1.5 and thus follows its license. Llama 2 is licensed under the LLAMA 2 Community License, Copyright (c) Meta Platforms, Inc. All Rights Reserved.

Intended use

Primary intended uses: The primary use of Ada LLaVA is research on multimodal large multimodal models and chatbots, especially for resource-constrain inference and deployment.

Primary intended users: The primary intended users of the model are researchers and hobbyists in computer vision, natural language processing, machine learning, and artificial intelligence.

Training dataset

• 665K image level instruction data from LLaVA-1.5 stage-2, see details in original LLaVA repo.

Limitations

While Ada-LLaVA is currently limited to processing one image at a time and only applies adaptive operations in its later half of layers, future work could explore multi-image input support and extend the adaptive mechanisms throughout the entire model architecture, including the vision encoder. These improvements would make the model more versatile and applicable to a broader range of real-world scenarios.