app.py

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from llama_cpp import Llama
from concurrent.futures import ThreadPoolExecutor, as_completed
from tqdm import tqdm
import uvicorn
from dotenv import load_dotenv
from difflib import SequenceMatcher
import re
from spaces import GPU
import httpx

# Cargar variables de entorno
load_dotenv()

# Inicializar aplicación FastAPI
app = FastAPI()

# Diccionario global para almacenar los modelos
global_data = {
    'models': []
}

# Configuración de los modelos (incluyendo los nuevos)
model_configs = [
    {"repo_id": "Ffftdtd5dtft/gpt2-xl-Q2_K-GGUF", "filename": "gpt2-xl-q2_k.gguf", "name": "GPT-2 XL"},
    {"repo_id": "Ffftdtd5dtft/Meta-Llama-3.1-8B-Instruct-Q2_K-GGUF", "filename": "meta-llama-3.1-8b-instruct-q2_k.gguf", "name": "Meta Llama 3.1-8B Instruct"},
    # Otros modelos omitidos por espacio
    {"repo_id": "Ffftdtd5dtft/Meta-Llama-3.1-70B-Instruct-Q2_K-GGUF", "filename": "meta-llama-3.1-70b-instruct-q2_k.gguf", "name": "Meta Llama 3.1-70B Instruct"},
    {"repo_id": "Ffftdtd5dtft/codegemma-2b-IQ1_S-GGUF", "filename": "codegemma-2b-iq1_s-imat.gguf", "name": "Codegemma 2B"},
    {"repo_id": "Ffftdtd5dtft/Mistral-Nemo-Instruct-2407-Q2_K-GGUF", "filename": "mistral-nemo-instruct-2407-q2_k.gguf", "name": "Mistral Nemo Instruct 2407"}
]

# Clase para gestionar modelos
class ModelManager:
    def __init__(self):
        self.models = []
    
    def load_model(self, model_config):
        print(f"Cargando modelo: {model_config['name']}...")
        return {"model": Llama.from_pretrained(repo_id=model_config['repo_id'], filename=model_config['filename']), "name": model_config['name']}
    
    @GPU(duration=0)
    def load_all_models(self):
        print("Iniciando carga de modelos...")
        with ThreadPoolExecutor(max_workers=len(model_configs)) as executor:
            futures = [executor.submit(self.load_model, config) for config in model_configs]
            models = []
            for future in tqdm(as_completed(futures), total=len(model_configs), desc="Cargando modelos", unit="modelo"):
                try:
                    model = future.result()
                    models.append(model)
                    print(f"Modelo cargado exitosamente: {model['name']}")
                except Exception as e:
                    print(f"Error al cargar el modelo: {e}")
        print("Todos los modelos han sido cargados.")
        return models

# Instanciar ModelManager y cargar modelos una sola vez
model_manager = ModelManager()
global_data['models'] = model_manager.load_all_models()

# Modelo global para la solicitud de chat
class ChatRequest(BaseModel):
    message: str
    top_k: int = 50
    top_p: float = 0.95
    temperature: float = 0.7

# Función para generar respuestas de chat
def generate_chat_response(request, model_data):
    try:
        user_input = normalize_input(request.message)
        llm = model_data['model']
        response = llm.create_chat_completion(
            messages=[{"role": "user", "content": user_input}],
            top_k=request.top_k,
            top_p=request.top_p,
            temperature=request.temperature
        )
        reply = response['choices'][0]['message']['content']
        return {"response": reply, "literal": user_input, "model_name": model_data['name']}
    except Exception as e:
        return {"response": f"Error: {str(e)}", "literal": user_input, "model_name": model_data['name']}

def normalize_input(input_text):
    return input_text.strip()

def remove_duplicates(text):
    text = re.sub(r'(Hello there, how are you\? \[/INST\]){2,}', 'Hello there, how are you? [/INST]', text)
    text = re.sub(r'(How are you\? \[/INST\]){2,}', 'How are you? [/INST]', text)
    text = text.replace('[/INST]', '')
    lines = text.split('\n')
    unique_lines = list(dict.fromkeys(lines))
    return '\n'.join(unique_lines).strip()

def remove_repetitive_responses(responses):
    seen = set()
    unique_responses = []
    for response in responses:
        normalized_response = remove_duplicates(response['response'])
        if normalized_response not in seen:
            seen.add(normalized_response)
            unique_responses.append(response)
    return unique_responses

# Manejo de errores en la inicialización de modelos (traza mencionada en el error)
def handle_initialization_error(allow_token):
    try:
        client = httpx.Client()
        pid = 0  # Variable que simula el proceso actual
        assert client.allow(allow_token=allow_token, pid=pid) == httpx.codes.OK
    except AssertionError:
        raise HTTPException(status_code=500, detail="Error en la inicialización del cliente Spaces")

# Ruta para generar chat en múltiples modelos
@app.post("/chat/")
async def chat(request: ChatRequest):
    try:
        # Simulación del error `AssertionError` durante la inicialización
        allow_token = "test_token"
        handle_initialization_error(allow_token)

        with ThreadPoolExecutor() as executor:
            futures = [executor.submit(generate_chat_response, request, model) for model in global_data['models']]
            responses = [future.result() for future in as_completed(futures)]
        unique_responses = remove_repetitive_responses(responses)
        return {"responses": unique_responses}
    except Exception as e:
        raise HTTPException(status_code=500, detail=f"Error procesando la solicitud: {str(e)}")

# Uso de template `chat_template.default`
chat_template = """
User: {message}
Bot: {response}
"""

# Plantilla de respuesta de chat
def render_chat_template(message, response):
    return chat_template.format(message=message, response=response)

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8000)