handler.py

from typing import Dict, List, Any
from transformers import pipeline
from PIL import Image    
import requests

import torch
from transformers import AutoProcessor, LlavaForConditionalGeneration

class EndpointHandler():
    def __init__(self, path="."):
        self.model = LlavaForConditionalGeneration.from_pretrained(
            path, 
            torch_dtype=torch.float16, 
            low_cpu_mem_usage=True, 
        ).to(0)
        self.processor = AutoProcessor.from_pretrained(path)

    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        """
        data args:
            inputs (:obj: `str`)
            date (:obj: `str`)
        Return:
            A :obj:`list` | `dict`: will be serialized and returned
        """
        # get inputs
        prompt = "USER: <image>\nWhat's in the image\nASSISTANT:"
        default_url = "https://cdn.faire.com/fastly/3c335e5c06d3027964ee8351093784c94dfa264e5eb26430c803f4ab3c44da84.jpeg"
        url = data.pop("image_url", default_url)
        inputs = data.pop("inputs", None)

        image = Image.open(requests.get(url, stream=True).raw)


        inputs = self.processor(prompt, image, return_tensors='pt').to(0, torch.float16)

        # run normal prediction
        output = self.model.generate(**inputs, max_new_tokens=200, do_sample=False)
        print(output)
        
        return output