handler.py

import torch.nn.functional as F
from torch import Tensor
from transformers import AutoTokenizer, AutoModel, AutoProcessor
from torch import cuda

class EndpointHandler():
    def __init__(self, path=""):
        self.processor = AutoProcessor.from_pretrained(path)
        self.model = AutoModel.from_pretrained(path, trust_remote_code=True)
        self.device = "cuda" if cuda.is_available() else "cpu"
        self.model.to(self.device)

    def __call__(self, data: Dict[str, Any]) -> List[List[int]]:
        image  = data.pop("inputs",data)

        processed = self.processor(images=image, return_tensors="pt").to(self.device)

        prediction = self.model(processed["pixel_values"])
        
        return prediction.item()