Update app.py

app.py

@@ -5,6 +5,7 @@ from concurrent.futures import ThreadPoolExecutor, as_completed
 import uvicorn
 import re
 from dotenv import load_dotenv
+from spaces import GPU
 
 load_dotenv()
 
@@ -98,6 +99,7 @@ def remove_duplicates(text):
             seen_lines.add(line)
     return '\n'.join(unique_lines)
 
+@GPU(duration=0)
 def generate_model_response(model, inputs, top_k, top_p, temperature):
     try:
         response = model.generate(inputs, top_k=top_k, top_p=top_p, temperature=temperature)