Reduce VRAM consumption by swapping `cuda()` and `to(torch.bfloat16)`

When I test the code locally, it appears that converting the weights to bfloat16 only after moving to the GPU causes the excess VRAM to not be freed up (unless maybe `torch.cuda.empty_cache()` is used, but this is simpler).

app.py

@@ -43,7 +43,7 @@ def process_ocr_task(image, model_size, task_type, ref_text):
         return "Please upload an image first.", None
 
     print("🚀 Moving model to GPU...")
-    model_gpu = model.cuda().to(torch.bfloat16)
+    model_gpu = model.to(torch.bfloat16).cuda()
     print("✅ Model is on GPU.")
 
     with tempfile.TemporaryDirectory() as output_path: