import gradio as gr
import torch
from PIL import Image
from transformers import MllamaForConditionalGeneration, AutoProcessor
from transformers import TextStreamer
from torchvision.transforms import Resize

# Define the model and processor
model_id = "0llheaven/Llama-3.2-11B-Vision-Radiology-mini"

device = "cuda" if torch.cuda.is_available() else "cpu"
model = MllamaForConditionalGeneration.from_pretrained(
    model_id,
    load_in_4bit=True,
    torch_dtype=torch.bfloat16,
    device_map=device,
)

model.gradient_checkpointing_enable()

processor = AutoProcessor.from_pretrained(model_id)

# Function to process the image and generate the description
def generate_description(image: Image.Image, instruction: str):
    image = image.convert("RGB")
    # image = Resize((224, 224))(image)

    # Create the message to pass to the model
    instruction = "You are an expert radiographer. Describe accurately what you see in this image."
    messages = [
        {"role": "user", "content": [
            {"type": "image"},
            {"type": "text", "text": instruction}
        ]}
    ]

    input_text = processor.apply_chat_template(messages, add_generation_prompt=True)
    inputs = processor(
        image,
        input_text,
        add_special_tokens=False,
        return_tensors="pt"
    ).to(model.device)

    # Generate the output from the model
    output = model.generate(**inputs, max_new_tokens=256)
    return processor.decode(output[0])

# Define Gradio interface
interface = gr.Interface(
    fn=generate_description,
    inputs=gr.Image(type="pil", label="Upload an Image"),
    outputs=gr.Textbox(label="Generated Description"),
    live=True,
    title="Radiology Image Description Generator",
    description="Upload an image and provide an instruction to generate a description using a vision-language model."
)

# Launch the interface
interface.launch()