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

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 memoria
llms = [Llama.from_pretrained(repo_id=model['repo_id'], filename=model['filename']) 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 = 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 reply
    except Exception as e:
        return f"Error: {str(e)}"

def select_best_response(responses, request):
    coherent_responses = filter_by_coherence(responses, request)
    best_response = filter_by_similarity(coherent_responses)
    return best_response

def filter_by_coherence(responses, request):
    # 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):
    # Ejecutar en ThreadPoolExecutor sin límite explícito de workers
    with ThreadPoolExecutor(max_workers=None) as executor:
        futures = [executor.submit(generate_chat_response, request, llm) for llm in llms]
        responses = []
        for future in as_completed(futures):
            response = future.result()
            responses.append(response)
    
    if any("Error" in response for response in responses):
        error_response = next(response for response in responses if "Error" in response)
        raise HTTPException(status_code=500, detail=error_response)
    
    best_response = select_best_response(responses, request)
    return {"response": best_response}

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