Load model on CPU during download and move to CUDA if available

app.py

@@ -184,11 +184,10 @@ with gr.Blocks() as demo:
             elif model_id.startswith("Qwen/Qwen3-VL"):
                 model_loader = Qwen3VLForConditionalGeneration
             assert model_loader is not None, f"Unsupported model ID: {model_id}"
+            # Load model on CPU to avoid using CUDA resources during download
             current_model = model_loader.from_pretrained(
-                model_id,
-                torch_dtype="auto",
-                device_map="auto",
-            ).eval()
+                model_id, torch_dtype=torch.bfloat16, device_map="cpu"
+            )
             current_processor = AutoProcessor.from_pretrained(model_id)
             current_model_id = model_id
         return current_model, current_processor
@@ -205,6 +204,10 @@ with gr.Blocks() as demo:
         image_resize: str,
         image_target_size: int | None,
     ):
+        # Move model to CUDA if available (inside @spaces.GPU decorated function)
+        model = model.to(DEVICE)
+        model.eval()
+
         base64_image = image_to_base64(
             resize_image(image, image_target_size)
             if image_resize == "Yes" and image_target_size