OPEA/llava-v1.5-7b-int4-sym-inc

Model Details

This model is an int4 model with group_size 128 and symmetric quantization of liuhaotian/llava-v1.5-7b generated by intel/auto-round. Load the model with revision="a121fa0" to use AutoGPTQ format.

How To Use

Requirements

1. Clone this repository and navigate to LLaVA folder

git clone https://github.com/haotian-liu/LLaVA.git
cd LLaVA

2. Refine LLaVA repo

vi llava/model/language_model/llava_llama.py
# add 'cache_position = None,' to line 71.

3. Install Package

pip install --upgrade pip  # enable PEP 660 support
pip install -e .

INT4 Inference

from auto_round import AutoRoundConfig ## must import for auto-round format
import requests
import torch
from PIL import Image
from llava.model.builder import load_pretrained_model
from llava.train.train import preprocess, preprocess_multimodal, DataCollatorForSupervisedDataset
class DataArgs:
    is_multimodal = True
    mm_use_im_start_end = False

quantized_model_path="OPEA/llava-v1.5-7b-int4-sym-inc"

tokenizer, model, image_processor, _ = load_pretrained_model(
    quantized_model_path,
    model_base=None,
    model_name=quantized_model_path,
    torch_dtype="auto",
    device_map="auto",
    ##revision="a121fa0" ##AutoGPTQ format
)
image_url = "http://images.cocodataset.org/train2017/000000116003.jpg"
messages = [{"from": "human", "value": "What is the tennis player doing in the image?\n<image>"}]

# Preparation for inference
image = Image.open(requests.get(image_url, stream=True).raw).convert('RGB')
image_input = image_processor.preprocess(image, return_tensors='pt')['pixel_values'][0].to(model.device)
input_data = preprocess_multimodal([messages], DataArgs())
inputs = preprocess(input_data, tokenizer, has_image=(image_input is not None))

output = model.generate(inputs['input_ids'].to(model.device), images=image_input.unsqueeze(0).half(), max_new_tokens=50)
print(tokenizer.batch_decode(output))

##INT4: The tennis player is celebrating a victory, raising his arms in the air, and holding his tennis racket.

##BF16: The tennis player is celebrating a victory, raising his arms in the air, and holding a tennis racket.

image_url = "http://images.cocodataset.org/train2017/000000411975.jpg"
messages = [{"from": "human", "value": "How many people are on the baseball field in the picture?\n<image>"}]

##INT4: There are three people on the baseball field in the picture.

##BF16: There are three people on the baseball field in the picture.


image_url = "http://images.cocodataset.org/train2017/000000093025.jpg"
messages = [{"from": "human", "value": "How many people and animals are there in the image?\n<image>"}]

##INT4: There are two people and one animal in the image.

##BF16: There are two people and one animal in the image.

Evaluation the model

pip3 install lmms_eval. The evaluation process may encounter errors that require changing model backend or evaluation code. Detailed instructions will be provided in a future update

auto-round-mllm --lmms --model OPEA/llava-v1.5-7b-int4-sym-inc --tasks pope,textvqa_val,scienceqa,mmbench_en  --output_dir "./eval_result" --device cuda:0 

Metric	16bits	Pile Calib INT4	Llava Calib INT4
avg	65.40	65.91	65.79
MMBench_DEV_EN_V11	64.09	64.43	64.43
ScienceQA_VAL	64.87	67.20	66.80
TextVQA_VAL	45.56	45.71	45.81
POPE	87.09	86.31	86.12

Generate the model

Here is the sample command to reproduce the model.

pip install auto-round
auto-round-mllm \
--model liuhaotian/llava-v1.5-7b \
--device 0 \
--group_size 128 \
--bits 4 \
--iters 1000 \
--nsample 512 \
--seqlen 2048 \
--format 'auto_gptq,auto_round' \
--output_dir "./tmp_autoround"

Ethical Considerations and Limitations

The model can produce factually incorrect output, and should not be relied on to produce factually accurate information. Because of the limitations of the pretrained model and the finetuning datasets, it is possible that this model could generate lewd, biased or otherwise offensive outputs.

Therefore, before deploying any applications of the model, developers should perform safety testing.

Caveats and Recommendations

Users (both direct and downstream) should be made aware of the risks, biases and limitations of the model.

Here are a couple of useful links to learn more about Intel's AI software:

• Intel Neural Compressor link

Disclaimer

The license on this model does not constitute legal advice. We are not responsible for the actions of third parties who use this model. Please consult an attorney before using this model for commercial purposes.

Cite

@article{cheng2023optimize, title={Optimize weight rounding via signed gradient descent for the quantization of llms}, author={Cheng, Wenhua and Zhang, Weiwei and Shen, Haihao and Cai, Yiyang and He, Xin and Lv, Kaokao and Liu, Yi}, journal={arXiv preprint arXiv:2309.05516}, year={2023} }

arxiv github