app.py

from fastapi import FastAPI, HTTPException
from pydantic import BaseModel
from llama_cpp import Llama
from concurrent.futures import ProcessPoolExecutor, as_completed
import uvicorn
from dotenv import load_dotenv
from difflib import SequenceMatcher
from tqdm import tqdm
import threading

load_dotenv()

app = FastAPI()

# Configuración de los modelos
models = [
    {"repo_id": "Ffftdtd5dtft/gpt2-xl-Q2_K-GGUF", "filename": "gpt2-xl-q2_k.gguf"},
    {"repo_id": "Ffftdtd5dtft/Meta-Llama-3.1-8B-Instruct-Q2_K-GGUF", "filename": "meta-llama-3.1-8b-instruct-q2_k.gguf"},
    {"repo_id": "Ffftdtd5dtft/gemma-2-9b-it-Q2_K-GGUF", "filename": "gemma-2-9b-it-q2_k.gguf"},
    {"repo_id": "Ffftdtd5dtft/gemma-2-27b-Q2_K-GGUF", "filename": "gemma-2-27b-q2_k.gguf"},
]

# Función para cargar un modelo
def load_model(model_config):
    return Llama.from_pretrained(repo_id=model_config['repo_id'], filename=model_config['filename'])

# Cargar modelos en paralelo
def load_all_models():
    with ProcessPoolExecutor() as executor:
        future_to_model = {executor.submit(load_model, model): model for model in models}
        loaded_models = {}
        for future in as_completed(future_to_model):
            model = future_to_model[future]
            try:
                loaded_models[model['repo_id']] = future.result()
                print(f"Modelo cargado en RAM: {model['repo_id']}")
            except Exception as exc:
                print(f"Error al cargar modelo {model['repo_id']}: {exc}")
    return loaded_models

# Cargar modelos en memoria
llms = load_all_models()

class ChatRequest(BaseModel):
    message: str
    top_k: int = 50
    top_p: float = 0.95
    temperature: float = 0.7

# Función global para generar respuestas de chat
def generate_chat_response(request, llm):
    try:
        user_input = normalize_input(request.message)
        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}
    except Exception as e:
        return {"response": f"Error: {str(e)}", "literal": user_input}

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

def filter_duplicates(responses):
    seen = set()
    unique_responses = []
    for response in responses:
        lines = response.split('\n')
        unique_lines = set()
        for line in lines:
            if line not in seen:
                seen.add(line)
                unique_lines.add(line)
        unique_responses.append('\n'.join(unique_lines))
    return unique_responses

def select_best_response(responses):
    # Eliminar respuestas duplicadas
    unique_responses = filter_duplicates(responses)
    # Deduplicar respuestas
    unique_responses = list(set(unique_responses))
    # Filtrar respuestas coherentes
    coherent_responses = filter_by_coherence(unique_responses)
    # Seleccionar la mejor respuesta
    best_response = filter_by_similarity(coherent_responses)
    return best_response

def filter_by_coherence(responses):
    # Implementa aquí un filtro de coherencia si es necesario
    return responses

def filter_by_similarity(responses):
    responses.sort(key=len, reverse=True)
    best_response = responses[0]
    for i in range(1, len(responses)):
        ratio = SequenceMatcher(None, best_response, responses[i]).ratio()
        if ratio < 0.9:
            best_response = responses[i]
            break
    return best_response

def worker_function(llm, request):
    return generate_chat_response(request, llm)

@app.post("/generate_chat")
async def generate_chat(request: ChatRequest):
    if not request.message.strip():
        raise HTTPException(status_code=400, detail="The message cannot be empty.")
    
    print(f"Procesando solicitud: {request.message}")

    responses = []
    threads = []

    # Crear un hilo para cada modelo
    for llm in llms.values():
        thread = threading.Thread(target=lambda: responses.append(worker_function(llm, request)))
        threads.append(thread)
        thread.start()

    # Esperar a que todos los hilos terminen
    for thread in threads:
        thread.join()

    # Seleccionar la mejor respuesta
    best_response = select_best_response([response['response'] for response in responses])
    
    print(f"Mejor respuesta seleccionada: {best_response}")

    return {
        "best_response": best_response,
        "all_responses": [response['response'] for response in responses]
    }

if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=7860)