---
license: apache-2.0
---
# Intruduction
We introduce Xmodel-VLM, a cutting-edge multimodal vision language model. It is designed for efficient deployment on consumer GPU servers. 
Our work directly confronts a pivotal industry issue by grappling with the prohibitive service costs that hinder the broad adoption of large-scale multimodal systems.

Refer to [our paper](https://arxiv.org/pdf/2405.09215) and [github](https://github.com/XiaoduoAILab/XmodelVLM) for more details!

To use Xmodel_VLM for the inference, all you need to do is to input a few lines of codes as demonstrated below. **However, please make sure that you are using the latest code and related virtual environments.**

## Inference example
```
import sys
import torch
import argparse
from PIL import Image
from pathlib import Path
import time
sys.path.append(str(Path(__file__).parent.parent.resolve()))

from xmodelvlm.model.xmodelvlm import load_pretrained_model
from xmodelvlm.conversation import conv_templates, SeparatorStyle
from xmodelvlm.utils import disable_torch_init, process_images, tokenizer_image_token, KeywordsStoppingCriteria
from xmodelvlm.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN

def inference_once(args):
    disable_torch_init()
    model_name = args.model_path.split('/')[-1]
    tokenizer, model, image_processor, context_len = load_pretrained_model(args.model_path, args.load_8bit, args.load_4bit)

    images = [Image.open(args.image_file).convert("RGB")]
    images_tensor = process_images(images, image_processor, model.config).to(model.device, dtype=torch.float16)

    conv = conv_templates[args.conv_mode].copy()
    conv.append_message(conv.roles[0], DEFAULT_IMAGE_TOKEN + "\n" + args.prompt)
    conv.append_message(conv.roles[1], None)
    prompt = conv.get_prompt()
    stop_str = conv.sep if conv.sep_style != SeparatorStyle.TWO else conv.sep2
    # Input
    input_ids = (tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors="pt").unsqueeze(0).cuda())
    stopping_criteria = KeywordsStoppingCriteria([stop_str], tokenizer, input_ids)
    # Inference
    with torch.inference_mode():
        start_time = time.time()
        output_ids = model.generate(
            input_ids,
            images=images_tensor,
            do_sample=True if args.temperature > 0 else False,
            temperature=args.temperature,
            top_p=args.top_p,
            num_beams=args.num_beams,
            max_new_tokens=args.max_new_tokens,
            use_cache=True,
            stopping_criteria=[stopping_criteria],
        )
        end_time = time.time()
        execution_time = end_time-start_time
        print("the execution time (secend): ", execution_time)
    # Result-Decode
    input_token_len = input_ids.shape[1]
    n_diff_input_output = (input_ids != output_ids[:, :input_token_len]).sum().item()
    if n_diff_input_output > 0:
        print(f"[Warning] {n_diff_input_output} output_ids are not the same as the input_ids")
    outputs = tokenizer.batch_decode(output_ids[:, input_token_len:], skip_special_tokens=True)[0]
    outputs = outputs.strip()
    if outputs.endswith(stop_str):
        outputs = outputs[: -len(stop_str)]
    print(f"🚀 {model_name}: {outputs.strip()}\n")
if __name__ == '__main__':
   model_path = "XiaoduoAILab/Xmodel_VLM" # model weight file
   image_file = "assets/demo.jpg" # image file
   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)
```

<center>
  
Prompt: **Who is the author of this book?\nAnswer the question using a single word or phrase.**
![Book Cover](https://huggingface.co/XiaoduoAILab/Xmodel_VLM/resolve/main/demo.jpg)
Author: **Susan Wise Bauer**

</center>

## Evaluation

We evaluate the multimodal performance across a variety of datasets: **VizWiz**, **SQA<sup>I</sup>**, **VQA<sup>T</sup>**, **POPE**, **GQA**, **MMB**, **MMB<sup>CN</sup>**
, **MM-Vet**, and **MME**. Our analysis, as depicted In the following table.

|     Method     |        LLM       | Res. | VizWiz |  SQA |  VQA | POPE |  GQA |  MMB | MMB<sup>CN</sup> | MM-Vet |   MME  |
|:--------------:|:----------------:|:----:|:------:|:----:|:----:|:----:|:----:|:----:|:--------:|:------:|:------:|
|  Openflamingo  |      MPT-7B      |  336 |    -   |   -  | 33.6 |   -  |   -  |  4.6 |     -    |    -   |    -   |
|     BLIP-2     |    Vicuna-13B    |  224 |    -   | 61.0 | 42.5 | 85.3 | 41.0 |   -  |     -    |    -   | 1293.8 |
|    MiniGPT-4   |     Vicuna-7B    |  224 |    -   |   -  |   -  |   -  | 32.2 | 23.0 |     -    |    -   |  581.7 |
|  InstructBLIP  |     Vicuna-7B    |  224 |    -   | 60.5 | 50.1 |   -  | 49.2 |   -  |     -    |    -   |    -   |
|  InstructBLIP  |    Vicuna-13B    |  224 |    -   | 63.1 | 50.7 | 78.9 | 49.5 |   -  |     -    |    -   | 1212.8 |
|     Shikra     |    Vicuna-13B    |  224 |    -   |   -  |   -  |   -  |   -  | 58.8 |     -    |    -   |    -   |
|     Qwen-VL    |      Qwen-7B     |  448 |    -   | 67.1 | 63.8 |   -  | 59.3 | 38.2 |     -    |    -   | 1487.6 |
|   MiniGPT-v2   |     LLaMA-7B     |  448 |    -   |   -  |   -  |   -  | 60.3 | 12.2 |     -    |    -   |    -   |
| LLaVA-v1.5-13B |    Vicuna-13B    |  336 |  53.6  | 71.6 | 61.3 | 85.9 | 63.3 | 67.7 |   63.6   |  35.4  | 1531.3 |
|  MobileVLM 1.7 | MobileLLaMA 1.4B |  336 |  26.3  | 54.7 | 41.5 | 84.5 | 56.1 | 53.2 |   16.67  |  21.7  | 1196.2 |
|   **Xmodel-VLM**   |  **Xmodel-LM 1.1B**  |  **336** |  **41.7**  | **53.3** | **39.9** | **85.9** | **58.3** | **52.0** |   **45.7**   |  **21.8**  | **1250.7** |