handler.py

from typing import Any, Dict, List
from transformers import Idefics2Processor, Idefics2ForConditionalGeneration
import torch
import logging
from PIL import Image
import requests


class EndpointHandler:
    def __init__(self, path=""):
        # Preload all the elements you are going to need at inference.
        self.logger = logging.getLogger()
        self.logger.addHandler(logging.StreamHandler())
        self.device = "cuda" if torch.cuda.is_available() else "cpu"
        self.processor = Idefics2Processor.from_pretrained(path)
        self.model = Idefics2ForConditionalGeneration.from_pretrained(path)
        self.model.to(self.device)
        self.logger.info("Initialisation finished!")

    def __call__(self, data: Dict[str, Any]) -> List[Dict[str, Any]]:
        """
         data args:
              inputs (:obj: `str` | `PIL.Image` | `np.array`)
              kwargs
        Return:
              A :obj:`list` | `dict`: will be serialized and returned
        """
        """image = data.pop("inputs", data)
        self.logger.info("image")

        # process image
        inputs = self.processor(images=image, return_tensors="pt").to(self.device)
        self.logger.info("inputs")
        self.logger.info(f"{inputs.input_ids}")
        generated_ids = self.model.generate(**inputs)
        self.logger.info("generated")

        # run prediction
        generated_text = self.processor.batch_decode(
            generated_ids, skip_special_tokens=True
        )
        self.logger.info("decoded")"""

        url_2 = "http://images.cocodataset.org/val2017/000000219578.jpg"

        image_1 = data.pop("inputs", data)
        image_2 = Image.open(requests.get(url_2, stream=True).raw)
        images = [image_1, image_2]

        messages = [
            {
                "role": "user",
                "content": [
                    {
                        "type": "text",
                        "text": "What’s the difference between these two images?",
                    },
                    {"type": "image"},
                    {"type": "image"},
                ],
            }
        ]

        self.model.to(self.device)

        # at inference time, one needs to pass `add_generation_prompt=True` in order to make sure the model completes the prompt
        text = self.processor.apply_chat_template(messages, add_generation_prompt=True)
        self.logger.info(text)
        # 'User: What’s the difference between these two images?<image><image><end_of_utterance>\nAssistant:'

        inputs = self.processor(images=images, text=text, return_tensors="pt").to(
            self.device
        )
        self.logger.info("inputs")

        generated_text = self.model.generate(**inputs, max_new_tokens=500)
        self.logger.info("generated")
        generated_text = self.processor.batch_decode(
            generated_text, skip_special_tokens=True
        )[0]
        self.logger.info(f"Generated text: {generated_text}")

        # decode output
        return generated_text