model_loader.py

import os
import threading
import torch
from transformers import AutoProcessor, AutoModelForVision2Seq

MODEL_NAME = "Qwen/Qwen2.5-VL-7B-Instruct"

model = None
processor = None
device = "cuda" if torch.cuda.is_available() else "cpu"

_model_lock = threading.Lock()


def get_model():
    global model, processor, device

    if model is None or processor is None:
        with _model_lock:
            if model is None or processor is None:
                print("Loading processor...")
                processor = AutoProcessor.from_pretrained(
                    MODEL_NAME,
                    trust_remote_code=True,
                    use_fast=True,
                    min_pixels=224 * 224,   # add on 8/5/26
                    max_pixels=1536 * 1536  # add on 8/5/26
                )

                print("Loading model...")
                model = AutoModelForVision2Seq.from_pretrained(
                    MODEL_NAME,
                    trust_remote_code=True,
                    torch_dtype=torch.float16,
                    device_map="auto",
                    low_cpu_mem_usage=True
                )

                model.eval()
                print("Model loaded successfully")

    return model, processor, device