Update app.py

app.py

@@ -1,7 +1,7 @@
 from fastapi import FastAPI, HTTPException
 from pydantic import BaseModel
 from llama_cpp import Llama
-from concurrent.futures import ProcessPoolExecutor
+from concurrent.futures import ThreadPoolExecutor
 import uvicorn
 from dotenv import load_dotenv
 from difflib import SequenceMatcher
@@ -10,6 +10,7 @@ load_dotenv()
 
 app = FastAPI()
 
+# Inicialización 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"},
@@ -17,10 +18,8 @@ models = [
     {"repo_id": "Ffftdtd5dtft/gemma-2-27b-Q2_K-GGUF", "filename": "gemma-2-27b-q2_k.gguf"},
 ]
 
-llms = []
-for model in models:
-    llm = Llama.from_pretrained(repo_id=model['repo_id'], filename=model['filename'])
-    llms.append(llm)
+# 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
@@ -48,6 +47,7 @@ def select_best_response(responses, request):
     return best_response
 
 def filter_by_coherence(responses, request):
+    # Puedes implementar un filtro más sofisticado si es necesario
     return responses
 
 def filter_by_similarity(responses):
@@ -62,7 +62,8 @@ def filter_by_similarity(responses):
 
 @app.post("/generate_chat")
 async def generate_chat(request: ChatRequest):
-    with ProcessPoolExecutor() as executor:
+    with ThreadPoolExecutor() as executor:
+        # Ejecutar las tareas en paralelo
         futures = [executor.submit(generate_chat_response, request, llm) for llm in llms]
         responses = [future.result() for future in futures]
     
@@ -70,6 +71,7 @@ async def generate_chat(request: ChatRequest):
         error_response = next(response for response in responses if "Error" in response)
         raise HTTPException(status_code=500, detail=error_response)
     
+    # Seleccionar la mejor respuesta
     best_response = select_best_response(responses, request)
     return {"response": best_response}