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 threading load_dotenv() app = FastAPI() # Configuración de los modelos model_configs = [ {"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/Phi-3-mini-128k-instruct-Q2_K-GGUF", "filename": "phi-3-mini-128k-instruct-q2_k.gguf"}, {"repo_id": "Ffftdtd5dtft/Meta-Llama-3.1-8B-Q2_K-GGUF", "filename": "meta-llama-3.1-8b-q2_k.gguf"}, {"repo_id": "Ffftdtd5dtft/starcoder2-3b-Q2_K-GGUF", "filename": "starcoder2-3b-q2_k.gguf"}, {"repo_id": "Ffftdtd5dtft/Qwen2-1.5B-Instruct-Q2_K-GGUF", "filename": "qwen2-1.5b-instruct-q2_k.gguf"}, ] def load_model(model_config): print(f"Cargando modelo {model_config['repo_id']}...") return Llama.from_pretrained(repo_id=model_config['repo_id'], filename=model_config['filename']) def load_all_models(): print("Iniciando carga de modelos...") with ThreadPoolExecutor(max_workers=len(model_configs)) as executor: futures = [executor.submit(load_model, config) for config in model_configs] models = [] for future in tqdm(as_completed(futures), total=len(model_configs), desc="Cargando modelos", unit="modelo"): try: model = future.result() models.append(model) print(f"Modelo cargado exitosamente: {model_configs[len(models)-1]['repo_id']}") except Exception as e: print(f"Error al cargar el modelo: {e}") print("Todos los modelos han sido cargados.") return models llms = load_all_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 filter_duplicates(responses): seen = set() unique_responses = [] for response in responses: lines = response.split('\n') unique_lines = set() for line in lines: if line not in seen: seen.add(line) unique_lines.add(line) unique_responses.append('\n'.join(unique_lines)) return unique_responses def select_best_response(responses): print("Filtrando respuestas...") unique_responses = filter_duplicates(responses) unique_responses = list(set(unique_responses)) coherent_responses = filter_by_coherence(unique_responses) 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 def worker_function(llm, request, progress_bar): print(f"Generando respuesta con el modelo...") response = generate_chat_response(request, llm) progress_bar.update(1) return 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}") responses = [] num_models = len(llms) with tqdm(total=num_models, desc="Generando respuestas", unit="modelo") as progress_bar: with ThreadPoolExecutor(max_workers=num_models) as executor: futures = [executor.submit(worker_function, llm, request, progress_bar) for llm in llms] for future in as_completed(futures): try: response = future.result() responses.append(response['response']) except Exception as exc: print(f"Error en la generación de respuesta: {exc}") best_response = select_best_response(responses) print(f"Mejor respuesta seleccionada: {best_response}") return { "best_response": best_response, "all_responses": responses } if __name__ == "__main__": uvicorn.run(app, host="0.0.0.0", port=7860)