MobileVLM: A Fast, Strong and Open
Vision Language Assistant for Mobile Devices

We present MobileVLM, a competent multimodal vision language model (MMVLM) targeted to run on mobile devices. It is an amalgamation of a myriad of architectural designs and techniques that are mobile-oriented, which comprises a set of language models at the scale of 1.4B and 2.7B parameters, trained from scratch, a multimodal vision model that is pre-trained in the CLIP fashion, cross-modality interaction via an efficient projector. We evaluate MobileVLM on several typical VLM benchmarks. Our models demonstrate on par performance compared with a few much larger models. More importantly, we measure the inference speed on both a Qualcomm Snapdragon 888 CPU and an NVIDIA Jeston Orin GPU, and we obtain state-of-the-art performance of 21.5 tokens and 65.3 tokens per second, respectively.

The MobileVLM architecture (right) utilizes MobileLLaMA as its language model, intakes $\mathbf{X}_v$ and $\mathbf{X}_q$ which are image and language instructions as respective inputs and gives $\mathbf{Y}_a$ as the output language response. LDP refers to a lightweight downsample projector (left).

📸 Release

• ⏳ MobileLLaMA Pre-training code.
• ⏳ MobileLLaMA SFT training code.
• Jan. 11st, 2024: The training and evaluation codes of MobileVLM are available now! Follow these step-by-step instructions below to easily train your own mobileVLM in 5 hours ⚡️ !
• Dec. 31st, 2023: Our MobileVLM weights are uploaded on the HuggingFace website. We also provide inference examples for the MobileLLaMA/MobileVLM model so that anyone can enjoy them early.
• Dec. 29th, 2023: Our MobileLLaMA weights are uploaded on the HuggingFace website. Enjoy them !
• Dec. 28th, 2023: 🔥🔥🔥 We release MobileVLM: A Fast, Strong and Open Vision Language Assistant for Mobile Devices on arxiv. Refer to our paper for more details !

🦙 Model Zoo

• MobileLLaMA-1.4B-Base
• MobileLLaMA-1.4B-Chat
• MobileLLaMA-2.7B-Base
• MobileLLaMA-2.7B-Chat
• MobileVLM-1.7B
• MobileVLM-3B

🔔 Usage and License Notices: 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. This project is licensed permissively under the Apache 2.0 license and does not impose any additional constraints. ^LLaVA

🛠️ Install

1. Clone this repository and navigate to MobileVLM folder

   git clone https://github.com/Meituan-AutoML/MobileVLM.git
   cd MobileVLM
   

2. Install Package

   conda create -n mobilevlm python=3.10 -y
   conda activate mobilevlm
   pip install --upgrade pip
   pip install -r requirements.txt
   

🗝️ Quick Start

Example for MobileLLaMA model inference

import torch
from transformers import LlamaTokenizer, LlamaForCausalLM

model_path = 'mtgv/MobileLLaMA-1.4B-Chat'

tokenizer = LlamaTokenizer.from_pretrained(model_path)
model = LlamaForCausalLM.from_pretrained(
    model_path, torch_dtype=torch.float16, device_map='auto',
)

prompt = 'Q: What is the largest animal?\nA:'
input_ids = tokenizer(prompt, return_tensors="pt").input_ids

generation_output = model.generate(
    input_ids=input_ids, max_new_tokens=32
)
print(tokenizer.decode(generation_output[0]))

• For more advanced usage, please follow the transformers LLaMA documentation.

Example for MobileVLM model inference

from scripts.inference import inference_once

model_path = "mtgv/MobileVLM-1.7B"
image_file = "assets/samples/demo.jpg"
prompt_str = "Who is the author of this book?\nAnswer the question using a single word or phrase."
# (or) What is the title of this book?
# (or) Is this book related to Education & Teaching?

args = type('Args', (), {
    "model_path": model_path,
    "image_file": image_file,
    "prompt": prompt_str,
    "conv_mode": "v1",
    "temperature": 0, 
    "top_p": None,
    "num_beams": 1,
    "max_new_tokens": 512,
    "load_8bit": False,
    "load_4bit": False,
})()

inference_once(args)

🪜 Step-by-step Tutorial

MobileVLM

The training process of MobileVLM is divided into two stages:

• stage I: feature alignment pretrain
  • ❄️ frozen vision encoder + 🔥 learnable LDP projector + ❄️ frozen LLM
  • this training process takes around 1~1.5 hours for MobileVLM-1.7B/3B on 8x A100 (80G) with a batch size of 256 and an average of approximately 15G/19G of GPU memory required.
• stage II: visual instruction tuning
  • ❄️ frozen vision encoder + 🔥 learnable LDP projector + 🔥 learnable LLM
  • this training process takes around 2~3.5 hours for MobileVLM-1.7B/3B on 8x A100 (80G) with a batch size of 128 and an average of approximately 46G/52G of GPU memory required.

Note: To train on fewer GPU memory or cards, you can reduce the per_device_train_batch_size and increase the gradient_accumulation_steps accordingly. Always keep the global batch size the same: per_device_train_batch_size x gradient_accumulation_steps x num_gpus.

1️⃣ Prepare MobileLLaMA checkpoints

Download MobileLLaMA chatbot checkpoints from huggingface website (🤗 1.7B, 2.7B). Please note that this is optional (it depends on your working environment), run the training script we provide below and the model will be automatically downloaded by the transformers library.

2️⃣ Prepare data

• For convenience, assume your working directory /path/to/project/mobilevlm as work_dir:

  • cd ${work_dir} && mkdir -p data/pretrain_data data/finetune_data data/benchmark_data

• prepare alignment pre-training data

  • cd ${work_dir}/data/pretrain_data
  • download the LLaVA-558K from here, which is provided by LLaVA team.

• prepare instruction tuning data

  • cd ${work_dir}/data/finetune_data
  • download the annotation of the LLaVA mixture instruction tuning data here, and download the images from constituting datasets: COCO, GQA, OCR-VQA, TextVQA, VisualGnome (Part1, Part2)

• prepare benchmark data

  • We evaluate models on a diverse set of 6 benchmarks, i.e. GQA, MMBench, MME, POPE, SQA, TextVQA. We do not evaluate using beam search to make the inference process consistent with the chat demo of real-time outputs. You should follow these instructions to manage the datasets.

  • Data Download Instructions

    • download some useful data/scripts pre-collected by us.
      • unzip benchmark_data.zip && cd benchmark_data
      • bmk_dir=${work_dir}/data/benchmark_data
    • gqa
      • download its image data following the official instructions here
      • cd ${bmk_dir}/gqa && ln -s /path/to/gqa/images images
    • mme
      • download the data following the official instructions here.
      • cd ${bmk_dir}/mme && ln -s /path/to/MME/MME_Benchmark_release_version images
    • pope
      • download coco from POPE following the official instructions here.
      • cd ${bmk_dir}/pope && ln -s /path/to/pope/coco coco && ln -s /path/to/coco/val2014 val2014
    • sqa
      • download images from the data/scienceqa folder of the ScienceQA repo.
      • cd ${bmk_dir}/sqa && ln -s /path/to/sqa/images images
    • textvqa
      • download images following the instructions here.
      • cd ${bmk_dir}/textvqa && ln -s /path/to/textvqa/train_images train_images
    • mmbench
      • no action is needed.

• organize the data directory as follows after downloading all of them:

  • Data Structure Tree

    .
    ├── benchmark_data
    │   ├── gqa
    │   │   ├── convert_gqa_for_eval.py
    │   │   ├── eval.py
    │   │   ├── images -> /path/to/your/gqa/images
    │   │   ├── llava_gqa_testdev_balanced.jsonl
    │   │   └── testdev_balanced_questions.json
    │   ├── mmbench
    │   │   ├── convert_mmbench_for_submission.py
    │   │   ├── eval.py
    │   │   └── mmbench_dev_en_20231003.tsv
    │   ├── mme
    │   │   ├── calculation.py
    │   │   ├── convert_answer_to_mme.py
    │   │   ├── images -> /path/to/your/MME/MME_Benchmark_release_version
    │   │   └── llava_mme.jsonl
    │   ├── pope
    │   │   ├── coco -> /path/to/your/pope/coco
    │   │   ├── eval.py
    │   │   ├── llava_pope_test.jsonl
    │   │   └── val2014 -> /path/to/your/coco/val2014
    │   ├── sqa
    │   │   ├── eval.py
    │   │   ├── images -> /path/to/your/scienceqa/images
    │   │   ├── llava_test_CQM-A.json
    │   │   ├── pid_splits.json
    │   │   └── problems.json
    │   └── textvqa
    │       ├── eval.py
    │       ├── llava_textvqa_val_v051_ocr.jsonl
    │       ├── m4c_evaluator.py
    │       ├── TextVQA_0.5.1_val.json
    │       └── train_images -> /path/to/your/textvqa/train_images
    ├── finetune_data
    │    ├── llava_v1_5_mix665k.json
    │    ├── coco
    │    │   └── train2017
    │    ├── gqa
    │    │   └── images
    │    ├── ocr_vqa
    │    │   └── images
    │    ├── textvqa
    │    │   └── train_images
    │    └── vg
    │        ├── VG_100K
    │        └── VG_100K_2
    ├── pretrain_data
    │    ├── images
    │    └── blip_laion_cc_sbu_558k.json
    

3️⃣ Run everything with one click!

LANGUAGE_MODEL=/path/to/your/MobileLLaMA-1.4B-Chat  # or 2.7B
VISION_MODEL=/path/to/your/clip-vit-large-patch14-336
bash run.sh mobilevlm1.7b pretrain-finetune-test ${LANGUAGE_MODEL} ${VISION_MODEL}

# (test-only) bash run.sh mobilevlm1.7b test /path/to/your/own/checkpoint
# (3B) bash run.sh mobilevlm3b pretrain-finetune-test ${LANGUAGE_MODEL} ${VISION_MODEL}

• Note 🧭: We place all running commands in run.sh so they can be run with one click for simplification. If you would like to modify some super-parameters to observe their impact, please dive into run.sh to explore.

🤝 Acknowledgments

• LLaVA: the codebase we built upon. Thanks for their wonderful work! 👏
• Vicuna: the amazing open-sourced large language model!

✏️ Reference

If you find MobileVLM or MobileLLaMA useful in your research or applications, please consider giving a star ⭐ and citing using the following BibTeX:

@misc{chu2023mobilevlm,
      title={MobileVLM: A Fast, Strong and Open Vision Language Assistant for Mobile Devices}, 
      author={Xiangxiang Chu and Limeng Qiao and Xinyang Lin and Shuang Xu and Yang Yang and Yiming Hu and Fei Wei and Xinyu Zhang and Bo Zhang and Xiaolin Wei and Chunhua Shen},
      year={2023},
      eprint={2312.16886},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

🌟 Star History