make run_inference to request GPUs

app.py

@@ -204,6 +204,7 @@ def prepare_for_inference():
     )
 
 
+@spaces.GPU
 def run_inference(word, image, mapping):
     """Main inference function, now focused just on generation"""
     if not word or not image or not mapping:
@@ -230,7 +231,6 @@ def run_inference(word, image, mapping):
     )
 
     # Start generation in a separate thread with torch.no_grad()
-    @spaces.GPU
     def generate_with_no_grad():
         with torch.no_grad():
             model.generate(**generation_kwargs)