Paper arxiv.org/abs/2409.14713

Two steps only need.

First step. (git clone and install required packages)

# Download Project Code
git clone https://github.com/ByungKwanLee/Phantom

# Virtual Environment
conda create -n trol python=3.11 -y
conda activate trol

# install torch
pip3 install torch torchvision

# install requiresments
pip install -r requirements.txt

# flash attention
pip install flash-attn --no-build-isolation

# all cache deleted 
conda clean -a && pip cache purge

Second step. (open, edit, and run demo.py)

# model selection
size = '7b' # [Select One] '0.5b' (transformers more recent version) | '1.8b' | '3.8b' (transformers==4.37.2) | '7b'

# User prompt
prompt_type="with_image" # Select one option "text_only", "with_image"
img_path='figures/demo.png'
question="Describe the image in detail"

# loading model
model, tokenizer = load_model(size=size)

# prompt type -> input prompt
if prompt_type == 'with_image':
    # Image Load
    image = pil_to_tensor(Image.open(img_path).convert("RGB"))
    inputs = [{'image': image, 'question': question}]
elif prompt_type=='text_only':
    inputs = [{'question': question}]

# cpu -> gpu
for param in model.parameters():
    if not param.is_cuda:
        param.data = param.cuda()

# Generate
with torch.inference_mode():

    # Model
    _inputs = model.eval_process(inputs=inputs,
                                data='demo',
                                tokenizer=tokenizer,
                                device='cuda:0')
    generate_ids = model.generate(**_inputs, do_sample=False, max_new_tokens=256)
answer = tokenizer.batch_decode(generate_ids, skip_special_tokens=True)[0]
print(answer)

So easy to run the code Let's shout Phantom!