app.py

from flask import Flask, request, jsonify, Response
import os
import logging
import threading
import time
import subprocess
import sys
subprocess.check_call([sys.executable, "-m", "pip", "install", "llama-cpp-python==0.3.16"])
from llama_cpp import Llama
import requests
import tempfile
import json
import gc
from concurrent.futures import ThreadPoolExecutor

app = Flask(__name__)
logging.basicConfig(level=logging.INFO)

MAX_CONTEXT_TOKENS = 1024 * 8
MAX_GENERATION_TOKENS = 1024 * 4

with open('engines.json', 'r') as f:
    MODELS = json.load(f)

class LLMManager:
    def __init__(self, models_config):
        self.models = {}
        self.models_config = models_config
        self.executor = ThreadPoolExecutor(max_workers=2)
        self.generation_lock = threading.Lock()
        self.session = requests.Session()
        adapter = requests.adapters.HTTPAdapter(pool_connections=2, pool_maxsize=2)
        self.session.mount('http://', adapter)
        self.session.mount('https://', adapter)
        self.load_all_models()

    def load_all_models(self):
        for model_config in self.models_config:
            try:
                model_name = model_config["name"]
                logging.info(f"🚀 Cargando modelo: {model_name}")
                
                temp_path = self._download_model(model_config["url"])
                
                actual_size = os.path.getsize(temp_path)
                actual_gb = actual_size / (1024*1024*1024)
                logging.info(f"📊 Tamaño descargado para {model_name}: {actual_gb:.2f} GB")

                n_batch = model_config.get("n_batch", 96)

                logging.info(f"🔄 Cargando {model_name} en RAM…")
                llm_instance = Llama(
                    model_path=temp_path,
                    n_ctx=MAX_CONTEXT_TOKENS,
                    n_batch=n_batch,
                    n_threads=2,
                    n_threads_batch=2,
                    use_mlock=False,
                    mmap=True,
                    low_vram=False,
                    vocab_only=False,
                    verbose=False,
                    logits_all=False,
                    mul_mat_q=True
                )
                
                os.remove(temp_path)

                self.models[model_name] = {
                    "instance": llm_instance,
                    "loaded": True,
                    "config": model_config
                }
                logging.info(f"✅ Modelo {model_name} cargado")

            except Exception as e:
                logging.error(f"❌ Error cargando modelo {model_config['name']}: {e}")
                self.models[model_config["name"]] = {
                    "instance": None,
                    "loaded": False,
                    "config": model_config,
                    "error": str(e)
                }

    def _download_model(self, model_url):
        temp_file = tempfile.NamedTemporaryFile(delete=False, suffix=".gguf")
        temp_path = temp_file.name
        temp_file.close()

        logging.info("📥 Descargando modelo…")
        
        response = self.session.get(model_url, stream=True, timeout=300)
        response.raise_for_status()

        downloaded = 0
        with open(temp_path, 'wb') as f:
            for chunk in response.iter_content(chunk_size=32768):
                if chunk:
                    f.write(chunk)
                    downloaded += len(chunk)
        
        return temp_path

    def get_model(self, model_name):
        return self.models.get(model_name)

    def chat_completion(self, model_name, messages, **kwargs):
        if not self.generation_lock.acquire(blocking=False):
            return {"error": "Servidor ocupado - Generación en progreso"}
        
        try:
            model_data = self.get_model(model_name)
            
            if not model_data or not model_data["loaded"]:
                error_msg = f"Modelo {model_name} no cargado"
                if model_data and "error" in model_data:
                    error_msg += f": {model_data['error']}"
                return {"error": error_msg}
            
            result = [None]
            exception = [None]
            
            def generate():
                try:
                    if 'repetition_penalty' in kwargs:
                        kwargs['repeat_penalty'] = kwargs.pop('repetition_penalty')
                    
                    result[0] = model_data["instance"].create_chat_completion(
                        messages=messages,
                        **kwargs
                    )
                except Exception as e:
                    exception[0] = e
            
            gen_thread = threading.Thread(target=generate, daemon=True)
            gen_thread.start()
            gen_thread.join(timeout=120)
            
            if gen_thread.is_alive():
                return {"error": "Timeout en generación (120 segundos)"}
            
            if exception[0]:
                raise exception[0]
            
            result[0]["provider"] = "telechars-ai"
            result[0]["model"] = model_name
            return result[0]
            
        finally:
            self.generation_lock.release()
            gc.collect()
    
    def get_loaded_models(self):
        loaded = []
        for name, data in self.models.items():
            if data["loaded"]:
                loaded.append(name)
        return loaded

    def get_all_models_status(self):
        status = {}
        for name, data in self.models.items():
            status[name] = {
                "loaded": data["loaded"],
                "url": data["config"]["url"]
            }
            if "error" in data:
                status[name]["error"] = data["error"]
        return status

llm_manager = LLMManager(MODELS)

@app.route('/')
def home():
    loaded_models = llm_manager.get_loaded_models()
    status_html = "<ul>"
    for model_name, model_data in llm_manager.models.items():
        status = "✅" if model_data["loaded"] else "❌"
        status_html += f"<li>{model_name}: {status}</li>"
    status_html += "</ul>"
    
    return f'''
    <!DOCTYPE html>
    <html>
    <head>
        <title>TeleChars AI API</title>
        <style>
            body {{ font-family: Arial, sans-serif; margin: 40px; }}
            .config {{ background: #f0f0f0; padding: 15px; border-radius: 5px; margin-bottom: 20px; }}
            .endpoint {{ background: #e8f4f8; padding: 10px; border-left: 4px solid #2196F3; margin: 10px 0; }}
        </style>
    </head>
    <body>
        <h1>TeleChars AI API</h1>
        
        <div class="config">
            <h3>⚙️ Configuración</h3>
            <p><strong>Max Context Tokens:</strong> {MAX_CONTEXT_TOKENS}</p>
            <p><strong>Max Generation Tokens:</strong> {MAX_GENERATION_TOKENS}</p>
        </div>
        
        <h2>📦 Modelos cargados:</h2>
        {status_html}
        <p>Total modelos: {len(loaded_models)}/{len(MODELS)}</p>
        
        <h2>🔗 Endpoints disponibles:</h2>
        <div class="endpoint">
            <strong>GET /generate/&lt;mensaje&gt;[?params]</strong><br>
            Devuelve solo el texto generado. Parámetros opcionales:<br>
            • system= (instrucciones del sistema)<br>
            • temperature= (0.0-2.0)<br>
            • top_p= (0.0-1.0)<br>
            • top_k= (0-100)<br>
            • model= (nombre del modelo)<br>
            • max_tokens= (máximo tokens a generar, default: {MAX_GENERATION_TOKENS})<br>
            • repetition_penalty= (penalización de repetición)<br>
            • presence_penalty= (penalización de presencia)<br>
            • frequency_penalty= (penalización de frecuencia)
        </div>
        
        <div class="endpoint">
            <strong>POST /v1/chat/completions</strong><br>
            Compatible con OpenAI API
        </div>
        
        <div class="endpoint">
            <strong>GET /health</strong><br>
            Estado del servicio
        </div>
        
        <div class="endpoint">
            <strong>GET /models</strong><br>
            Lista todos los modelos disponibles
        </div>
    </body>
    </html>
    '''

@app.route('/v1/chat/completions', methods=['POST'])
def chat_completions():
    try:
        data = request.get_json()
        messages = data.get('messages', [])
        model_name = data.get('model', MODELS[0]["name"])
        
        if model_name not in llm_manager.models:
            return jsonify({"error": f"Modelo '{model_name}' no encontrado. Modelos disponibles: {list(llm_manager.models.keys())}"}), 400
        
        kwargs = {}
        for key in data.keys():
            if key not in ['messages', 'model']:
                kwargs[key] = data[key]
        
        if 'max_tokens' not in kwargs:
            kwargs['max_tokens'] = MAX_GENERATION_TOKENS
        else:
            if kwargs['max_tokens'] > MAX_GENERATION_TOKENS:
                kwargs['max_tokens'] = MAX_GENERATION_TOKENS
        
        result = llm_manager.chat_completion(model_name, messages, **kwargs)

        if "error" in result:
            return jsonify(result), 500
            
        return jsonify(result), 200
        
    except Exception as e:
        return jsonify({"error": str(e)}), 500

@app.route('/generate/<path:user_message>', methods=['GET'])
def generate_endpoint(user_message):
    try:
        system_instruction = request.args.get('system', '')
        temperature = float(request.args.get('temperature', 0.7))
        top_p = float(request.args.get('top_p', 0.95))
        top_k = int(request.args.get('top_k', 0))
        model_name = request.args.get('model', MODELS[0]["name"])
        max_tokens = int(request.args.get('max_tokens', MAX_GENERATION_TOKENS))
        
        repetition_penalty = request.args.get('repetition_penalty')
        presence_penalty = request.args.get('presence_penalty')
        frequency_penalty = request.args.get('frequency_penalty')
        
        if not 0 <= temperature <= 2:
            return Response(
                f"Error: El parámetro 'temperature' debe estar entre 0 y 2",
                status=400,
                mimetype='text/plain'
            )
        
        if not 0 <= top_p <= 1:
            return Response(
                f"Error: El parámetro 'top_p' debe estar entre 0 y 1",
                status=400,
                mimetype='text/plain'
            )

        if not 0 <= top_k <= 100:
            return Response(
                f"Error: El parámetro 'top_k' debe estar entre 0 y 100",
                status=400,
                mimetype='text/plain'
            )
        
        if repetition_penalty:
            try:
                repetition_penalty = float(repetition_penalty)
            except ValueError:
                return Response(
                    "Error: repetition_penalty debe ser número válido",
                    status=400,
                    mimetype='text/plain'
                )
        
        if presence_penalty:
            try:
                presence_penalty = float(presence_penalty)
            except ValueError:
                return Response(
                    "Error: presence_penalty debe ser número válido",
                    status=400,
                    mimetype='text/plain'
                )
        
        if frequency_penalty:
            try:
                frequency_penalty = float(frequency_penalty)
            except ValueError:
                return Response(
                    "Error: frequency_penalty debe ser número válido",
                    status=400,
                    mimetype='text/plain'
                )
        
        if max_tokens > MAX_GENERATION_TOKENS:
            max_tokens = MAX_GENERATION_TOKENS
        
        if model_name not in llm_manager.models:
            return Response(
                f"Error: Modelo '{model_name}' no encontrado. Modelos disponibles: {', '.join(llm_manager.models.keys())}",
                status=400,
                mimetype='text/plain'
            )
        
        messages = [
            {"role": "system", "content": system_instruction},
            {"role": "user", "content": user_message}
        ]
        
        kwargs = {
            "temperature": temperature,
            "top_p": top_p,
            "max_tokens": max_tokens,
            "stream": False
        }

        if top_k:
            kwargs["top_k"] = int(top_k)
        
        if repetition_penalty:
            kwargs["repetition_penalty"] = repetition_penalty
        
        if presence_penalty:
            kwargs["presence_penalty"] = presence_penalty
            
        if frequency_penalty:
            kwargs["frequency_penalty"] = frequency_penalty
            
        result = llm_manager.chat_completion(model_name, messages, **kwargs)
        
        if "error" in result:
            return Response(
                f"Error: {result['error']}",
                status=500,
                mimetype='text/plain'
            )

        response_text = result.get("choices", [{}])[0].get("message", {}).get("content", "")
        
        if not response_text:
            response_text = "No se generó respuesta"
        
        return Response(
            response_text,
            status=200,
            mimetype='text/plain'
        )

    except ValueError as e:
        return Response(
            f"Error: Parámetros inválidos - {str(e)}. Asegúrate de que temperature, top_p y max_tokens sean números válidos.",
            status=400,
            mimetype='text/plain'
        )
    except Exception as e:
        return Response(
            f"Error: {str(e)}",
            status=500,
            mimetype='text/plain'
        )

@app.route('/health', methods=['GET'])
def health():
    loaded_models = llm_manager.get_loaded_models()
    return jsonify({
        "status": "healthy" if len(loaded_models) > 0 else "error",
        "loaded_models": loaded_models,
        "total_models": len(MODELS),
        "config": {
            "max_context_tokens": MAX_CONTEXT_TOKENS,
            "max_generation_tokens": MAX_GENERATION_TOKENS
        }
    })

@app.route('/models', methods=['GET'])
def list_models():
    return jsonify({
        "available_models": MODELS,
        "status": llm_manager.get_all_models_status(),
        "config": {
            "max_context_tokens": MAX_CONTEXT_TOKENS,
            "max_generation_tokens": MAX_GENERATION_TOKENS
        }
    })

@app.route('/models/<model_name>', methods=['GET'])
def get_model_status(model_name):
    model_data = llm_manager.get_model(model_name)
    if not model_data:
        return jsonify({"error": f"Modelo '{model_name}' no encontrado"}), 404
    
    return jsonify({
        "model": model_name,
        "loaded": model_data["loaded"],
        "url": model_data["config"]["url"],
        "error": model_data.get("error"),
        "config": {
            "max_context_tokens": MAX_CONTEXT_TOKENS,
            "max_generation_tokens": MAX_GENERATION_TOKENS
        }
    })

if __name__ == '__main__':
    app.run(host='0.0.0.0', port=7860, debug=False)