app/model_loader.py

from llama_cpp import Llama

llm = None  # Sẽ được khởi tạo sau

def load_model():
    global llm
    if llm is None:
        llm = Llama(
            model_path="models/tinyllama.gguf",
            n_ctx=2048,
            n_threads=4,
        )
    return llm