app.py

import gradio as gr

import os
import subprocess

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="bartowski/gemma-2-9b-it-GGUF", filename="gemma-2-9b-it-Q4_K_L.gguf", local_dir="./")

subprocess.run(["python3", "-m", "llama_cpp.server", "--model ./gemma-2-9b-it-Q4_K_L.gguf",  "--host localhost", "--port 8000", "--chat_format gemma"], shell=True)


with gr.Blocks() as demo:
    gr.Markdown("Running...")

demo.launch()