Fixes to Gradio App. Calling half when using CUDA devices.

app.py

@@ -25,10 +25,12 @@ model = DonutModel.from_pretrained(
                 ignore_mismatched_sizes=True,
             )
 
-device = "cuda" if torch.cuda.is_available() else "cpu"
+#device = "cuda" if torch.cuda.is_available() else "cpu"
+
+if has_cuda:
+    model.half()
+    model = model.to("cuda")
 
-model = model.to(device)
-model.eval()
 
 task_name = "matricula"
 task_prompt = f"<s_{task_name}>"
@@ -38,6 +40,8 @@ def predict_matricula(model, task_name, image):
     image = image.resize(size=(1280, 960), 
                          resample=Image.Resampling.NEAREST)
 
+    model.eval()
+    
     result = model.inference(image=image, prompt=f"<s_{task_name}>")["predictions"][0]
     return result