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
import logging
import os
import numpy as np
from functools import lru_cache
from cachetools import TTLCache
from multiprocessing import cpu_count
import threading
import queue

# Configuración de logging para suprimir mensajes de depuración innecesarios
logging.basicConfig(level=logging.ERROR)

# Cargar variables de entorno
load_dotenv()

# Inicializar aplicación FastAPI
app = FastAPI()

# Configuración de la caché
cache_size = 2000
cache_ttl = 7200
cache = TTLCache(maxsize=cache_size, ttl=cache_ttl)

# Diccionario global para almacenar los modelos en RAM
global_data = {
    'models': {}
}

# Configuración de los modelos
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"},
    {"repo_id": "Ffftdtd5dtft/gemma-2-9b-it-Q2_K-GGUF", "filename": "gemma-2-9b-it-q2_k.gguf", "name": "Gemma 2-9B IT"},
    {"repo_id": "Ffftdtd5dtft/gemma-2-27b-Q2_K-GGUF", "filename": "gemma-2-27b-q2_k.gguf", "name": "Gemma 2-27B"},
    {"repo_id": "Ffftdtd5dtft/Phi-3-mini-128k-instruct-Q2_K-GGUF", "filename": "phi-3-mini-128k-instruct-q2_k.gguf", "name": "Phi-3 Mini 128K Instruct"},
    {"repo_id": "Ffftdtd5dtft/Meta-Llama-3.1-8B-Q2_K-GGUF", "filename": "meta-llama-3.1-8b-q2_k.gguf", "name": "Meta Llama 3.1-8B"},
    {"repo_id": "Ffftdtd5dtft/Qwen2-7B-Instruct-Q2_K-GGUF", "filename": "qwen2-7b-instruct-q2_k.gguf", "name": "Qwen2 7B Instruct"},
    {"repo_id": "Ffftdtd5dtft/starcoder2-3b-Q2_K-GGUF", "filename": "starcoder2-3b-q2_k.gguf", "name": "Starcoder2 3B"},
    {"repo_id": "Ffftdtd5dtft/Qwen2-1.5B-Instruct-Q2_K-GGUF", "filename": "qwen2-1.5b-instruct-q2_k.gguf", "name": "Qwen2 1.5B Instruct"}
]

# Clase para gestionar modelos
class ModelManager:
    def __init__(self):
        self.models = {}
    
    def load_model(self, model_config):
        try:
            model = Llama.from_pretrained(repo_id=model_config['repo_id'], filename=model_config['filename'])
            self.models[model_config['name']] = model
            return model
        except Exception as e:
            logging.error(f"Error al cargar el modelo {model_config['name']}: {e}")
            return None
    
    def load_all_models(self):
        with ThreadPoolExecutor(max_workers=min(len(model_configs), cpu_count())) as executor:
            futures = [executor.submit(self.load_model, config) for config in model_configs]
            for future in tqdm(as_completed(futures), total=len(model_configs), desc="Cargando modelos", unit="modelo"):
                future.result()
        return self.models

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

# Clase 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
@lru_cache(maxsize=20000)
def generate_chat_response(request: ChatRequest, model_name: str):
    cache_key = f"{request.message}_{model_name}"
    
    if cache_key in cache:
        return cache[cache_key]
    
    model = global_data['models'].get(model_name)
    if not model:
        return {"response": "Error: Modelo no encontrado.", "literal": request.message, "model_name": model_name}
    
    try:
        user_input = normalize_input(request.message)
        response = model.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']
        
        # Almacenar en caché la respuesta
        cache[cache_key] = {"response": reply, "literal": user_input, "model_name": model_name}
        
        return cache[cache_key]
    except Exception as e:
        logging.error(f"Error en la generación de respuesta con el modelo {model_name}: {e}")
        return {"response": f"Error: {str(e)}", "literal": user_input, "model_name": model_name}

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

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

def select_best_response(responses):
    responses = remove_repetitive_responses(responses)
    responses = [remove_duplicates(response['response']) for response in responses]
    unique_responses = list(set(responses))
    coherent_responses = filter_by_coherence(unique_responses)
    best_response = filter_by_similarity(coherent_responses)
    return best_response

def filter_by_coherence(responses):
    responses.sort(key=len, reverse=True)
    return responses

def filter_by_similarity(responses):
    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(model_name, request, response_queue):
    try:
        response = generate_chat_response(request, model_name)
        response_queue.put((model_name, response))
    except Exception as e:
        logging.error(f"Error en la generación de respuesta con el modelo {model_name}: {e}")
        response_queue.put((model_name, {"response": f"Error: {str(e)}", "literal": request.message, "model_name": model_name}))

@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.")
    
    responses = []
    num_models = len(global_data['models'])
    response_queue = queue.Queue()

    with ThreadPoolExecutor(max_workers=min(num_models, cpu_count())) as executor:
        futures = [executor.submit(worker_function, model_name, request, response_queue) for model_name in global_data['models']]
        for future in tqdm(as_completed(futures), total=num_models, desc="Generando respuestas", unit="modelo"):
            future.result()
    
    while not response_queue.empty():
        model_name, response = response_queue.get()
        responses.append(response)
    
    best_response = select_best_response(responses)
    
    return {
        "best_response": best_response,
        "all_responses": responses
    }

# Cargar los modelos en la memoria RAM de manera más eficiente
def pre_load_models():
    for model_name, model in global_data['models'].items():
        model._load_model()  # Método hipotético para pre-cargar modelos en RAM

pre_load_models()

# Optimización de la carga de modelos en lotes
def optimize_model_loading():
    # Implementar carga de modelos en lotes con manejo eficiente de recursos
    batch_size = min(len(model_configs), cpu_count() * 2)
    for i in range(0, len(model_configs), batch_size):
        batch_configs = model_configs[i:i + batch_size]
        with ThreadPoolExecutor(max_workers=batch_size) as executor:
            futures = [executor.submit(model_manager.load_model, config) for config in batch_configs]
            for future in tqdm(as_completed(futures), total=len(batch_configs), desc="Optimizando carga de modelos", unit="modelo"):
                try:
                    model = future.result()
                    global_data['models'][batch_configs[futures.index(future)]['name']] = model
                except Exception as e:
                    logging.error(f"Error al optimizar la carga del modelo: {e}")

optimize_model_loading()

# Implementar técnicas de paralelización en la generación de respuestas
def parallelize_response_generation(request: ChatRequest):
    response_queue = queue.Queue()
    with ThreadPoolExecutor(max_workers=min(len(global_data['models']), cpu_count())) as executor:
        futures = [executor.submit(worker_function, model_name, request, response_queue) for model_name in global_data['models']]
        for future in tqdm(as_completed(futures), total=len(futures), desc="Generando respuestas en paralelo", unit="modelo"):
            future.result()
    
    responses = []
    while not response_queue.empty():
        responses.append(response_queue.get())
    return responses

@app.post("/generate_chat_parallel")
async def generate_chat_parallel(request: ChatRequest):
    if not request.message.strip():
        raise HTTPException(status_code=400, detail="The message cannot be empty.")
    
    responses = parallelize_response_generation(request)
    best_response = select_best_response(responses)
    
    return {
        "best_response": best_response,
        "all_responses": responses
    }

# Optimizar el uso de memoria
def optimize_memory_usage():
    import gc
    gc.collect()

# Ejecutar el servidor FastAPI
if __name__ == "__main__":
    uvicorn.run(app, host="0.0.0.0", port=8000)