app.py

import gradio as gr

# 🔒 Disable unnecessary API exposure by preventing auto-load
with gr.Blocks(fill_height=True) as demo:
    with gr.Sidebar():
        gr.Markdown("# Controlled Inference")
        gr.Markdown("This model operates under strict control. No external dependencies beyond the defined API.")
        button = gr.LoginButton("Sign in")

    # **Manual trigger required** for API call, preventing auto-loading of external inference
    load_button = gr.Button("Load Model")
    
    model_display = gr.Markdown("**Model Status: Unloaded**")

    def load_model():
        return "Model Loaded: meta-llama/Llama-3.3-70B-Instruct (Locked Mode)"

    load_button.click(load_model, outputs=model_display)
    
    gr.load("models/meta-llama/Llama-3.3-70B-Instruct", accept_token=button, provider="novita", visible=False)

demo.launch()