app.py

from fastapi import FastAPI, HTTPException, Request
from pydantic import BaseModel
import uvicorn
import requests
import asyncio
import os
import io
import time
from typing import List, Dict, Any
from llama_cpp import Llama  # Ajusta según la biblioteca que estés utilizando
from tqdm import tqdm

app = FastAPI()

# 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"},
    {"repo_id": "Ffftdtd5dtft/starcoder2-15b-Q2_K-GGUF", "filename": "starcoder2-15b-q2_k.gguf", "name": "Starcoder2 15B"},
    {"repo_id": "Ffftdtd5dtft/gemma-2-2b-it-Q2_K-GGUF", "filename": "gemma-2-2b-it-q2_k.gguf", "name": "Gemma 2-2B IT"},
    {"repo_id": "Ffftdtd5dtft/sarvam-2b-v0.5-Q2_K-GGUF", "filename": "sarvam-2b-v0.5-q2_k.gguf", "name": "Sarvam 2B v0.5"},
    {"repo_id": "Ffftdtd5dtft/WizardLM-13B-Uncensored-Q2_K-GGUF", "filename": "wizardlm-13b-uncensored-q2_k.gguf", "name": "WizardLM 13B Uncensored"},
    {"repo_id": "Ffftdtd5dtft/Qwen2-Math-72B-Instruct-Q2_K-GGUF", "filename": "qwen2-math-72b-instruct-q2_k.gguf", "name": "Qwen2 Math 72B Instruct"},
    {"repo_id": "Ffftdtd5dtft/WizardLM-7B-Uncensored-Q2_K-GGUF", "filename": "wizardlm-7b-uncensored-q2_k.gguf", "name": "WizardLM 7B Uncensored"},
    {"repo_id": "Ffftdtd5dtft/Qwen2-Math-7B-Instruct-Q2_K-GGUF", "filename": "qwen2-math-7b-instruct-q2_k.gguf", "name": "Qwen2 Math 7B Instruct"}
]

class ModelManager:
    def __init__(self):
        self.models = {}
        self.model_parts = {}
        self.load_lock = asyncio.Lock()
        self.index_lock = asyncio.Lock()
        self.part_size = 1024 * 1024  # Tamaño de cada parte en bytes (1 MB)
        self.max_loading_time = 0  # Tiempo máximo en segundos para cargar un modelo

    async def download_model_to_memory(self, model_config):
        url = f"https://huggingface.co/{model_config['repo_id']}/resolve/main/{model_config['filename']}"
        print(f"Descargando modelo desde {url}")
        try:
            response = requests.get(url)
            response.raise_for_status()
            print(f"Descarga completa para {model_config['name']}")
            return io.BytesIO(response.content)
        except requests.RequestException as e:
            raise HTTPException(status_code=500, detail=f"Error al descargar el modelo: {e}")

    async def save_model_to_temp_file(self, model_config):
        model_file = await self.download_model_to_memory(model_config)
        temp_filename = f"/tmp/{model_config['filename']}"
        print(f"Guardando el modelo en {temp_filename}")
        with open(temp_filename, 'wb') as f:
            f.write(model_file.getvalue())
        print(f"Modelo guardado en {temp_filename}")
        return temp_filename

    async def load_model(self, model_config):
        async with self.load_lock:
            try:
                start_time = time.time()
                temp_filename = await self.save_model_to_temp_file(model_config)
                elapsed_time = time.time() - start_time
                if elapsed_time > self.max_loading_time:
                    print(f"El modelo {model_config['name']} tardó {elapsed_time:.2f} segundos en cargar. Dividiendo el modelo.")
                    await self.handle_large_model(temp_filename, model_config)
                else:
                    print(f"Cargando modelo desde {temp_filename}")
                    llama = Llama(temp_filename)  # Ajusta según la biblioteca y clase correctas

                    tokenizer = llama.tokenizer
                    model_data = {
                        'model': llama,
                        'tokenizer': tokenizer,
                        'pad_token': tokenizer.pad_token,
                        'pad_token_id': tokenizer.pad_token_id,
                        'eos_token': tokenizer.eos_token,
                        'eos_token_id': tokenizer.eos_token_id,
                        'bos_token': tokenizer.bos_token,
                        'bos_token_id': tokenizer.bos_token_id,
                        'unk_token': tokenizer.unk_token,
                        'unk_token_id': tokenizer.unk_token_id
                    }

                    self.models[model_config['name']] = model_data
                    print(f"Modelo {model_config['name']} cargado correctamente")
            except Exception as e:
                print(f"Error al cargar el modelo: {e}")

    async def handle_large_model(self, model_filename, model_config):
        total_size = os.path.getsize(model_filename)
        num_parts = (total_size + self.part_size - 1) // self.part_size

        print(f"Modelo {model_config['name']} dividido en {num_parts} partes")
        with open(model_filename, 'rb') as file:
            for i in tqdm(range(num_parts), desc=f"Indexando {model_config['name']}"):
                start = i * self.part_size
                end = min(start + self.part_size, total_size)
                file.seek(start)
                model_part = io.BytesIO(file.read(end - start))
                await self.index_model_part(model_part, i)

    async def index_model_part(self, model_part, part_index):
        async with self.index_lock:
            part_name = f"part_{part_index}"
            print(f"Indexando parte {part_index}")
            llama_part = Llama(model_part)
            self.model_parts[part_name] = llama_part
            print(f"Parte {part_index} indexada")

    async def generate_response(self, user_input):
        print("Generando respuestas")
        tasks = [self.generate_chat_response(user_input, model_data) for model_data in self.models.values()]
        responses = await asyncio.gather(*tasks)
        return responses

    async def generate_chat_response(self, user_input, model_data):
        try:
            print(f"Generando respuesta usando el modelo {model_data['model']}")
            start_time = time.time()
            generated_text = model_data['model'].generate(user_input)
            elapsed_time = time.time() - start_time

            if len(generated_text) > 1000:
                parts = []
                while len(generated_text) > 1000:
                    part = generated_text[:1000]
                    parts.append(part)
                    generated_text = generated_text[1000:]
                parts.append(generated_text)
            else:
                parts = [generated_text]

            print(f"Respuesta generada usando el modelo {model_data['model']} en {elapsed_time:.2f} segundos")
            return {
                'model_name': model_data['model'],
                'generated_text_parts': parts
            }
        except Exception as e:
            print(f"Error al generar respuesta con el modelo {model_data['model']}: {e}")
            return {'model_name': model_data['model'], 'error': str(e)}

@app.post("/generate/")
async def generate(request: Request):
    data = await request.json()
    user_input = data.get('input', '')
    if not user_input:
        raise HTTPException(status_code=400, detail="Se requiere una entrada de usuario.")

    try:
        model_manager = ModelManager()
        tasks = [model_manager.load_model(config) for config in model_configs]
        await asyncio.gather(*tasks)
        responses = await model_manager.generate_response(user_input)
        return {"responses": responses}
    except Exception as e:
        raise HTTPException(status_code=500, detail=str(e))

def start_uvicorn():
    uvicorn.run(app, host="0.0.0.0", port=7860)

if __name__ == "__main__":
    loop = asyncio.get_event_loop()
    model_manager = ModelManager()
    tasks = [model_manager.load_model(config) for config in model_configs]
    loop.run_until_complete(asyncio.gather(*tasks))
    start_uvicorn()