app.py

import base64
import subprocess
from tempfile import NamedTemporaryFile

from fastapi import FastAPI, HTTPException
from PIL import Image
from pydantic import BaseModel
from transformers import AutoModelForCausalLM, AutoTokenizer
import uvicorn 

app = FastAPI()

# define request body
class RequestData(BaseModel):
    prompt: str
    image: str


def load_model():
    model_id = "vikhyatk/moondream2"
    revision = "2024-08-26"
    model = AutoModelForCausalLM.from_pretrained(
        model_id, trust_remote_code=True, revision=revision
    )
    tokenizer = AutoTokenizer.from_pretrained(model_id, revision=revision)
    return model, tokenizer


MODEL, TOKENIZER = load_model()
print("INFO: Model loaded successfully!")


@app.get("/")
def greet_json():
    return {"Hello": "World!"}


@app.post("/query")
def query(data: RequestData):
    prompt = data.prompt
    image = data.image

    try:
        # decode base64 to image
        image = base64.b64decode(image)

        with NamedTemporaryFile(delete=True, suffix=".png") as temp_image:
            temp_image.write(image)
            temp_image.flush()

            image = Image.open(temp_image.name)
            enc_image = MODEL.encode_image(image)
            response = MODEL.answer_question(enc_image, str(prompt), TOKENIZER)

            return {"response": str(response)}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

if __name__ == "__main__":
    uvicorn.run("app:app")