app.py

import json
import subprocess
import requests
import time
import socket
import gradio as gr

# Función para verificar si el servidor está activo en el puerto
def is_server_active(host, port):
    with socket.socket(socket.AF_INET, socket.SOCK_STREAM) as s:
        return s.connect_ex((host, port)) == 0

# Descarga y ejecución del modelo
url = "https://huggingface.co/mlabonne/NeuralBeagle14-7B-GGUF/resolve/main/neuralbeagle14-7b.Q4_K_M.gguf?download=true"
response = requests.get(url)
with open("./model.gguf", mode="wb") as file:
    file.write(response.content)
print("Model downloaded")

# Ejecutar el servidor LLM
command = ["python3", "-m", "llama_cpp.server", "--model", "./model.gguf", "--host", "0.0.0.0", "--port", "2600", "--n_threads", "2"]
server_process = subprocess.Popen(command)  # Almacenamos el proceso para poder terminarlo más tarde
print("Model server starting...")

# Esperar a que el servidor esté activo
while not is_server_active("0.0.0.0", 2600):
    print("Waiting for server to start...")
    time.sleep(5)
print("Model server is ready!")

def response(message, history):
    url = "http://localhost:2600/v1/completions"
    body = {"prompt": "[INST]"+message+"[/INST]", "max_tokens": 300, "echo": False, "stream": False}
    response_text = ""
    
    try:
        # Eliminado el timeout para esperar indefinidamente
        with requests.post(url, json=body, stream=True) as stream_response:
            for text_chunk in stream_response.iter_content(chunk_size=None):
                text = text_chunk.decode('utf-8')
                print("Respuesta cruda:", text)  # Imprimir la respuesta cruda para depuración

                if text.startswith("data: "):
                    text = text.replace("data: ", "")
                if text.startswith("{") and "choices" in text:
                    try:
                        response_json = json.loads(text)
                        part = response_json["choices"][0]["text"]
                        print(part, end="", flush=True)
                        response_text += part
                    except json.JSONDecodeError as e:
                        print("Error al decodificar JSON:", e)
                        break
                elif text.strip():
                    print("Respuesta no JSON:", text)
                    break
    except requests.exceptions.RequestException as e:
        print(f"Error al realizar la solicitud: {e}")

    yield response_text

def cleanup_server():
    print("Closing server...")
    server_process.terminate()  # Terminar el proceso del servidor
    server_process.wait()  # Esperar a que el proceso termine
    print("Server closed.")

# Configurar y lanzar la interfaz de Gradio
gr_interface = gr.ChatInterface(
    fn=response,
    title="NeuralBeagle 14-7b - By Maxime Labonne ❤️",
    theme='syddharth/gray-minimal'
)

try:
    gr_interface.launch(share=True)
finally:
    cleanup_server()  # Asegurarse de limpiar el servidor al finalizar