utils/model_serving/blip_serve.py

from dotenv import load_dotenv
from transformers import BlipForConditionalGeneration, BlipProcessor
import torch
import litserve as ls
import os

load_dotenv()

hf_token = os.getenv("HUGGINGFACE")

class RedionesBlipModel():
    def __init__(self):
        self.model_name = "Salesforce/blip-image-captioning-base"
        self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
        self.token = hf_token

    def setup(self, device):
        device = self.device
        self.model = BlipForConditionalGeneration.from_pretrained(self.model_name,
                                                                  use_auth_token=self.token,
                                                                  )
        self.tokenizer = BlipProcessor.from_pretrained(self.model_name, use_auth_token=self.token)
        self.model.to(device)
        self.model.eval()
        
    def predict(self, image):
        input_text = self.tokenizer(image, return_tensors="pt")
        outputs = self.model.generate(input_ids = input_text["input_ids"].to(self.device), max_new_tokens=50)
        
        return outputs