app.py

import spaces
import torch
import re
import gradio as gr
from PIL import Image
import io
from transformers import AutoTokenizer, AutoModelForCausalLM

if torch.cuda.is_available():
    device, dtype = "cuda", torch.float16
else:
    device, dtype = "cpu", torch.float32

model_id = "vikhyatk/moondream2"
revision = "2024-04-02"
tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
moondream = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, revision=revision
).to(device=device, dtype=dtype)
moondream.eval()

@spaces.GPU(duration=10)
def answer_questions(images, prompt_text):
    prompts = [p.strip() for p in prompt_text.split(',')]  # Splitting and cleaning prompts
    # Extracting images from tuples and converting to PIL images
    image_objects = [Image.open(io.BytesIO(img[0])) for img in images]
    image_embeds = [moondream.encode_image(img) for img in image_objects]
    answers = moondream.batch_answer(
        images=image_embeds,
        prompts=prompts,
        tokenizer=tokenizer,
    )
    return ["\n".join(ans) for ans in answers]

with gr.Blocks() as demo:
    gr.Markdown("# 🌔 moondream2\nA tiny vision language model. [GitHub](https://github.com/vikhyatk/moondream)")
    with gr.Row():
        img = gr.Gallery(label="Upload Images", type="pil")
        prompt = gr.Textbox(label="Input Prompts", placeholder="Enter prompts separated by commas. Ex: Describe this image, What is in this image?", lines=2)
        submit = gr.Button("Submit")
    output = gr.TextArea(label="Responses", lines=4)
    submit.click(answer_questions, [img, prompt], output)

demo.queue().launch()