README.md
| 1 | --- |
| 2 | language: |
| 3 | - en |
| 4 | datasets: |
| 5 | - liuhaotian/LLaVA-Instruct-150K |
| 6 | pipeline_tag: image-text-to-text |
| 7 | arxiv: 2304.08485 |
| 8 | license: llama2 |
| 9 | tags: |
| 10 | - vision |
| 11 | - image-text-to-text |
| 12 | --- |
| 13 | # LLaVA Model Card |
| 14 | |
| 15 |  |
| 16 | |
| 17 | Below is the model card of Llava model 7b, which is copied from the original Llava model card that you can find [here](https://huggingface.co/liuhaotian/llava-v1.5-13b). |
| 18 | |
| 19 | Check out also the Google Colab demo to run Llava on a free-tier Google Colab instance: [](https://colab.research.google.com/drive/1qsl6cd2c8gGtEW1xV5io7S8NHh-Cp1TV?usp=sharing) |
| 20 | |
| 21 | Or check out our Spaces demo! [](https://huggingface.co/spaces/llava-hf/llava-4bit) |
| 22 | |
| 23 | |
| 24 | ## Model details |
| 25 | |
| 26 | **Model type:** |
| 27 | LLaVA is an open-source chatbot trained by fine-tuning LLaMA/Vicuna on GPT-generated multimodal instruction-following data. |
| 28 | It is an auto-regressive language model, based on the transformer architecture. |
| 29 | |
| 30 | **Model date:** |
| 31 | LLaVA-v1.5-7B was trained in September 2023. |
| 32 | |
| 33 | **Paper or resources for more information:** |
| 34 | https://llava-vl.github.io/ |
| 35 | |
| 36 | ## How to use the model |
| 37 | |
| 38 | First, make sure to have `transformers >= 4.35.3`. |
| 39 | The model supports multi-image and multi-prompt generation. Meaning that you can pass multiple images in your prompt. Make sure also to follow the correct prompt template (`USER: xxx\nASSISTANT:`) and add the token `<image>` to the location where you want to query images: |
| 40 | |
| 41 | ### Using `pipeline`: |
| 42 | |
| 43 | Below we used [`"llava-hf/llava-1.5-7b-hf"`](https://huggingface.co/llava-hf/llava-1.5-7b-hf) checkpoint. |
| 44 | |
| 45 | ```python |
| 46 | from transformers import pipeline |
| 47 | |
| 48 | pipe = pipeline("image-text-to-text", model="llava-hf/llava-1.5-7b-hf") |
| 49 | messages = [ |
| 50 | { |
| 51 | "role": "user", |
| 52 | "content": [ |
| 53 | {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/ai2d-demo.jpg"}, |
| 54 | {"type": "text", "text": "What does the label 15 represent? (1) lava (2) core (3) tunnel (4) ash cloud"}, |
| 55 | ], |
| 56 | }, |
| 57 | ] |
| 58 | |
| 59 | out = pipe(text=messages, max_new_tokens=20) |
| 60 | print(out) |
| 61 | >>> [{'input_text': [{'role': 'user', 'content': [{'type': 'image', 'url': 'https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/transformers/tasks/ai2d-demo.jpg'}, {'type': 'text', 'text': 'What does the label 15 represent? (1) lava (2) core (3) tunnel (4) ash cloud'}]}], 'generated_text': 'Lava'}] |
| 62 | ``` |
| 63 | |
| 64 | ### Using pure `transformers`: |
| 65 | |
| 66 | Below is an example script to run generation in `float16` precision on a GPU device: |
| 67 | |
| 68 | ```python |
| 69 | import requests |
| 70 | from PIL import Image |
| 71 | |
| 72 | import torch |
| 73 | from transformers import AutoProcessor, LlavaForConditionalGeneration |
| 74 | |
| 75 | model_id = "llava-hf/llava-1.5-7b-hf" |
| 76 | model = LlavaForConditionalGeneration.from_pretrained( |
| 77 | model_id, |
| 78 | torch_dtype=torch.float16, |
| 79 | low_cpu_mem_usage=True, |
| 80 | ).to(0) |
| 81 | |
| 82 | processor = AutoProcessor.from_pretrained(model_id) |
| 83 | |
| 84 | # Define a chat history and use `apply_chat_template` to get correctly formatted prompt |
| 85 | # Each value in "content" has to be a list of dicts with types ("text", "image") |
| 86 | conversation = [ |
| 87 | { |
| 88 | |
| 89 | "role": "user", |
| 90 | "content": [ |
| 91 | {"type": "text", "text": "What are these?"}, |
| 92 | {"type": "image"}, |
| 93 | ], |
| 94 | }, |
| 95 | ] |
| 96 | prompt = processor.apply_chat_template(conversation, add_generation_prompt=True) |
| 97 | |
| 98 | image_file = "http://images.cocodataset.org/val2017/000000039769.jpg" |
| 99 | raw_image = Image.open(requests.get(image_file, stream=True).raw) |
| 100 | inputs = processor(images=raw_image, text=prompt, return_tensors='pt').to(0, torch.float16) |
| 101 | |
| 102 | output = model.generate(**inputs, max_new_tokens=200, do_sample=False) |
| 103 | print(processor.decode(output[0][2:], skip_special_tokens=True)) |
| 104 | ``` |
| 105 | |
| 106 | ----------- |
| 107 | From transformers>=v4.48, you can also pass image url or local path to the conversation history, and let the chat template handle the rest. |
| 108 | Chat template will load the image for you and return inputs in `torch.Tensor` which you can pass directly to `model.generate()` |
| 109 | |
| 110 | ```python |
| 111 | messages = [ |
| 112 | { |
| 113 | "role": "user", |
| 114 | "content": [ |
| 115 | {"type": "image", "url": "https://www.ilankelman.org/stopsigns/australia.jpg"} |
| 116 | {"type": "text", "text": "What is shown in this image?"}, |
| 117 | ], |
| 118 | }, |
| 119 | ] |
| 120 | |
| 121 | inputs = processor.apply_chat_template(messages, add_generation_prompt=True, tokenize=True, return_dict=True, return_tensors"pt") |
| 122 | output = model.generate(**inputs, max_new_tokens=50) |
| 123 | ``` |
| 124 | |
| 125 | ### Model optimization |
| 126 | |
| 127 | #### 4-bit quantization through `bitsandbytes` library |
| 128 | |
| 129 | First make sure to install `bitsandbytes`, `pip install bitsandbytes` and make sure to have access to a CUDA compatible GPU device. Simply change the snippet above with: |
| 130 | |
| 131 | ```diff |
| 132 | model = LlavaForConditionalGeneration.from_pretrained( |
| 133 | model_id, |
| 134 | torch_dtype=torch.float16, |
| 135 | low_cpu_mem_usage=True, |
| 136 | + load_in_4bit=True |
| 137 | ) |
| 138 | ``` |
| 139 | |
| 140 | #### Use Flash-Attention 2 to further speed-up generation |
| 141 | |
| 142 | First make sure to install `flash-attn`. Refer to the [original repository of Flash Attention](https://github.com/Dao-AILab/flash-attention) regarding that package installation. Simply change the snippet above with: |
| 143 | |
| 144 | ```diff |
| 145 | model = LlavaForConditionalGeneration.from_pretrained( |
| 146 | model_id, |
| 147 | torch_dtype=torch.float16, |
| 148 | low_cpu_mem_usage=True, |
| 149 | + use_flash_attention_2=True |
| 150 | ).to(0) |
| 151 | ``` |
| 152 | |
| 153 | ## License |
| 154 | Llama 2 is licensed under the LLAMA 2 Community License, |
| 155 | Copyright (c) Meta Platforms, Inc. All Rights Reserved. |