app.py

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

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"},
]

# Cargar modelos en RAM solo una vez
llms = [Llama.from_pretrained(repo_id=model['repo_id'], filename=model['filename']) for model in models]
print(f"Modelos cargados en RAM: {[model['repo_id'] for model in models]}")

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

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 select_best_response(responses):
    # Deduplicar respuestas
    unique_responses = list(set(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

@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}")

    # Utilizar un ThreadPoolExecutor para procesar los modelos en paralelo
    with ThreadPoolExecutor() as executor:
        futures = [executor.submit(generate_chat_response, request, llm) for llm in llms]
        responses = []

        for future in tqdm(as_completed(futures), total=len(futures), desc="Generando respuestas"):
            response = future.result()
            responses.append(response)
            print(f"Modelo procesado: {response['literal'][:30]}...")

    # Extraer respuestas de los diccionarios
    response_texts = [resp['response'] for resp in responses]
    
    # Verificar si hay errores en las respuestas
    error_responses = [resp for resp in responses if "Error" in resp['response']]
    if error_responses:
        error_response = error_responses[0]
        raise HTTPException(status_code=500, detail=error_response['response'])
    
    # Seleccionar la mejor respuesta
    best_response = select_best_response(response_texts)
    
    print(f"Mejor respuesta seleccionada: {best_response}")

    return {
        "best_response": best_response,
        "all_responses": response_texts
    }

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