Add gpu enabling

app.py

@@ -35,8 +35,13 @@ def initialize_models():
     global kosmos_model, kosmos_processor, zephyr_pipe, audio_model, audio_config
     try:
         print("Loading Kosmos-2...")
-        kosmos_model = AutoModelForVision2Seq.from_pretrained("microsoft/kosmos-2-patch14-224")
-        kosmos_processor = AutoProcessor.from_pretrained("microsoft/kosmos-2-patch14-224")
+        kosmos_model = AutoModelForVision2Seq.from_pretrained(
+            "microsoft/kosmos-2-patch14-224",
+            device_map="auto",
+            torch_dtype=torch.float16
+            )
+        kosmos_processor = AutoProcessor.from_pretrained(
+            "microsoft/kosmos-2-patch14-224")
         if torch.cuda.is_available():
             kosmos_model = kosmos_model.to("cuda")
     except Exception as e:
@@ -260,6 +265,7 @@ def get_image_examples():
     
     return examples
 
+@spaces.GPU(enable_queue=True)
 def infer(image_in, api_status):
     if image_in is None:
         raise gr.Error("Please provide an image of architecture")