---
datasets:
- Lin-Chen/ShareGPT4V
pipeline_tag: image-text-to-text
library_name: xtuner
---

<div align="center">
  <img src="https://github.com/InternLM/lmdeploy/assets/36994684/0cf8d00f-e86b-40ba-9b54-dc8f1bc6c8d8" width="600"/>


[![Generic badge](https://img.shields.io/badge/GitHub-%20XTuner-black.svg)](https://github.com/InternLM/xtuner)


</div>

## Model

llava-phi-3-mini is a LLaVA model fine-tuned from [microsoft/Phi-3-mini-4k-instruct](https://huggingface.co/microsoft/Phi-3-mini-4k-instruct) and [CLIP-ViT-Large-patch14-336](https://huggingface.co/openai/clip-vit-large-patch14-336) with [ShareGPT4V-PT](https://huggingface.co/datasets/Lin-Chen/ShareGPT4V) and [InternVL-SFT](https://github.com/OpenGVLab/InternVL/tree/main/internvl_chat#prepare-training-datasets) by [XTuner](https://github.com/InternLM/xtuner).


**Note: This model is in official LLaVA format.**

Resources:

- GitHub: [xtuner](https://github.com/InternLM/xtuner)
- HuggingFace LLaVA format model: [xtuner/llava-phi-3-mini-hf](https://huggingface.co/xtuner/llava-phi-3-mini-hf)
- GGUF LLaVA model: [xtuner/llava-phi-3-mini-gguf](https://huggingface.co/xtuner/llava-phi-3-mini-gguf)
- XTuner LLaVA format model: [xtuner/llava-phi-3-mini-xtuner](https://huggingface.co/xtuner/llava-phi-3-mini-xtuner)


## Details

| Model                 | Visual      Encoder | Projector | Resolution |   Pretraining Strategy | Fine-tuning      Strategy |      Pretrain     Dataset |    Fine-tune     Dataset | Pretrain Epoch | Fine-tune Epoch |
| :-------------------- | ------------------: | --------: | ---------: | ---------------------: | ------------------------: | ------------------------: | -----------------------: | -------------- | --------------- |
| LLaVA-v1.5-7B         |              CLIP-L |       MLP |        336 | Frozen LLM, Frozen ViT |      Full LLM, Frozen ViT |       LLaVA-PT     (558K) |     LLaVA-Mix     (665K) | 1              | 1               |
| LLaVA-Llama-3-8B      |              CLIP-L |       MLP |        336 | Frozen LLM, Frozen ViT |        Full LLM, LoRA ViT |       LLaVA-PT     (558K) |     LLaVA-Mix     (665K) | 1              | 1               |
| LLaVA-Llama-3-8B-v1.1 |              CLIP-L |       MLP |        336 | Frozen LLM, Frozen ViT |        Full LLM, LoRA ViT | ShareGPT4V-PT     (1246K) | InternVL-SFT     (1268K) | 1              | 1               |
| **LLaVA-Phi-3-mini**  |              CLIP-L |       MLP |        336 | Frozen LLM, Frozen ViT |        Full LLM, Full ViT | ShareGPT4V-PT     (1246K) | InternVL-SFT     (1268K) | 1              | 2               |

## Results

<div  align="center">
<img src="https://github.com/InternLM/xtuner/assets/36994684/78524f65-260d-4ae3-a687-03fc5a19dcbb" alt="Image" width=500" />
</div>

| Model                 | MMBench Test (EN) | MMMU  Val | SEED-IMG | AI2D Test | ScienceQA Test | HallusionBench aAcc | POPE | GQA  | TextVQA |   MME    | MMStar |
| :-------------------- | :---------------: | :-------: | :------: | :-------: | :------------: | :-----------------: | :--: | :--: | :-----: | :------: | :----: |
| LLaVA-v1.5-7B         |       66.5        |   35.3    |   60.5   |   54.8    |      70.4      |        44.9         | 85.9 | 62.0 |  58.2   | 1511/348 |  30.3  |
| LLaVA-Llama-3-8B      |       68.9        |   36.8    |   69.8   |   60.9    |      73.3      |        47.3         | 87.2 | 63.5 |  58.0   | 1506/295 |  38.2  |
| LLaVA-Llama-3-8B-v1.1 |       72.3        |   37.1    |   70.1   |   70.0    |      72.9      |        47.7         | 86.4 | 62.6 |  59.0   | 1469/349 |  45.1  |
| **LLaVA-Phi-3-mini**  |       69.2        |   41.4    |   70.0   |   69.3    |      73.7      |        49.8         | 87.3 | 61.5 |  57.8   | 1477/313 |  43.7  |



## Quickstart

### Chat by LLaVA official library

1. Install official LLaVA library

```bash
pip install git+https://github.com/haotian-liu/LLaVA.git
```

2. Chat by below script

<details> 
    <summary>cli.py</summary>

```python
import argparse
from io import BytesIO

import requests
import torch
from llava.constants import DEFAULT_IMAGE_TOKEN, IMAGE_TOKEN_INDEX
from llava.conversation import Conversation, SeparatorStyle
from llava.mm_utils import process_images, tokenizer_image_token
from llava.model import LlavaLlamaForCausalLM
from PIL import Image
from transformers import (AutoTokenizer, BitsAndBytesConfig, StoppingCriteria,
                          StoppingCriteriaList, TextStreamer)


def load_image(image_file):
    if image_file.startswith('http://') or image_file.startswith('https://'):
        response = requests.get(image_file)
        image = Image.open(BytesIO(response.content)).convert('RGB')
    else:
        image = Image.open(image_file).convert('RGB')
    return image


class StopWordStoppingCriteria(StoppingCriteria):
    """StopWord stopping criteria."""

    def __init__(self, tokenizer, stop_word):
        self.tokenizer = tokenizer
        self.stop_word = stop_word
        self.length = len(self.stop_word)

    def __call__(self, input_ids, *args, **kwargs) -> bool:
        cur_text = self.tokenizer.decode(input_ids[0])
        cur_text = cur_text.replace('\r', '').replace('\n', '')
        return cur_text[-self.length:] == self.stop_word


def get_stop_criteria(tokenizer, stop_words=[]):
    stop_criteria = StoppingCriteriaList()
    for word in stop_words:
        stop_criteria.append(StopWordStoppingCriteria(tokenizer, word))
    return stop_criteria


def main(args):
    kwargs = {'device_map': args.device}
    if args.load_8bit:
        kwargs['load_in_8bit'] = True
    elif args.load_4bit:
        kwargs['load_in_4bit'] = True
        kwargs['quantization_config'] = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_compute_dtype=torch.float16,
            bnb_4bit_use_double_quant=True,
            bnb_4bit_quant_type='nf4')
    else:
        kwargs['torch_dtype'] = torch.float16

    tokenizer = AutoTokenizer.from_pretrained(args.model_path)
    model = LlavaLlamaForCausalLM.from_pretrained(
        args.model_path, low_cpu_mem_usage=True, **kwargs)
    vision_tower = model.get_vision_tower()
    if not vision_tower.is_loaded:
        vision_tower.load_model(device_map=args.device)
    image_processor = vision_tower.image_processor

    conv = Conversation(
        system=system='<|system|>\nAnswer the questions.',
        roles=('<|user|>\n', '<|assistant|>\n'),
        messages=[],
        offset=0,
        sep_style=SeparatorStyle.MPT,
        sep='<|end|>',
    )
    roles = conv.roles

    image = load_image(args.image_file)
    image_size = image.size
    image_tensor = process_images([image], image_processor, model.config)

    if type(image_tensor) is list:
        image_tensor = [
            image.to(model.device, dtype=torch.float16)
            for image in image_tensor
        ]
    else:
        image_tensor = image_tensor.to(model.device, dtype=torch.float16)

    while True:
        try:
            inp = input(f'{roles[0]}: ')
        except EOFError:
            inp = ''
        if not inp:
            print('exit...')
            break

        print(f'{roles[1]}: ', end='')

        if image is not None:
            inp = DEFAULT_IMAGE_TOKEN + '\n' + inp
            image = None

        conv.append_message(conv.roles[0], inp)
        conv.append_message(conv.roles[1], None)
        prompt = conv.get_prompt()

        input_ids = tokenizer_image_token(
            prompt, tokenizer, IMAGE_TOKEN_INDEX,
            return_tensors='pt').unsqueeze(0).to(model.device)
        stop_criteria = get_stop_criteria(
            tokenizer=tokenizer, stop_words=[conv.sep])

        streamer = TextStreamer(
            tokenizer, skip_prompt=True, skip_special_tokens=True)

        with torch.inference_mode():
            output_ids = model.generate(
                input_ids,
                images=image_tensor,
                image_sizes=[image_size],
                do_sample=True if args.temperature > 0 else False,
                temperature=args.temperature,
                max_new_tokens=args.max_new_tokens,
                streamer=streamer,
                stopping_criteria=stop_criteria,
                use_cache=True)

        outputs = tokenizer.decode(output_ids[0]).strip()
        conv.messages[-1][-1] = outputs

        if args.debug:
            print('\n', {'prompt': prompt, 'outputs': outputs}, '\n')


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument(
        '--model-path', type=str, default='xtuner/llava-llama-3-8b-v1_1-hf')
    parser.add_argument('--image-file', type=str, required=True)
    parser.add_argument('--device', type=str, default='auto')
    parser.add_argument('--temperature', type=float, default=0.2)
    parser.add_argument('--max-new-tokens', type=int, default=512)
    parser.add_argument('--load-8bit', action='store_true')
    parser.add_argument('--load-4bit', action='store_true')
    parser.add_argument('--debug', action='store_true')
    args = parser.parse_args()
    main(args)
```

</details>

```
python ./cli.py  --model-path xtuner/llava-phi-3-mini --image-file https://raw.githubusercontent.com/open-mmlab/mmdeploy/main/tests/data/tiger.jpeg  --load-4bit
```


### Reproduce

Please refer to [docs](https://github.com/InternLM/xtuner/tree/main/xtuner/configs/llava/phi3_mini_4k_instruct_clip_vit_large_p14_336#readme).

## Citation

```bibtex
@misc{2023xtuner,
    title={XTuner: A Toolkit for Efficiently Fine-tuning LLM},
    author={XTuner Contributors},
    howpublished = {\url{https://github.com/InternLM/xtuner}},
    year={2023}
}
```