handler.py

from transformers import AutoProcessor, Blip2ForConditionalGeneration
import base64
from io import BytesIO
from PIL import Image
import string
import torch

class EndpointHandler:
    def __init__(self, path=""):
        # load model and processor from path
        self.processor = AutoProcessor.from_pretrained(path)
        self.model = Blip2ForConditionalGeneration.from_pretrained(path, device_map="auto", load_in_4bit=True)

    def __call__(self, data):
        """
        Args:
            inputs:
                Dict of image and text inputs.
        """
        # process input
        inputs = data.pop("inputs", data)
        image = Image.open(BytesIO(base64.b64decode(inputs['image'])))
        inputs = self.processor(images=image, text=inputs["text"], return_tensors="pt").to("cuda", torch.float16)
        generated_ids = self.model.generate(
            **inputs,
            temperature=1.0,
            length_penalty=1.0,
            repetition_penalty=1.5,
            max_length=30,
            min_length=1,
            num_beams=5,
            top_p=0.9,
        )
        result = self.processor.batch_decode(generated_ids, skip_special_tokens=True)[0].strip()
        if result and result[-1] not in string.punctuation:
            result += "."


        return [{"generated_text": result}]