predict_example.py

from pathlib import Path
import json
from transformers import AutoModel, AutoTokenizer

root = Path(__file__).resolve().parent
tokenizer = AutoTokenizer.from_pretrained(root, trust_remote_code=True, local_files_only=True)
model = AutoModel.from_pretrained(root, trust_remote_code=True, local_files_only=True).eval()
request = json.loads((root / "serving/example_request.json").read_text())
print(json.dumps(model.decide(tokenizer, **request), indent=2))