Should work.....

app.py

@@ -177,7 +177,7 @@ def preprocess_for_removal(images, masks):
         out_masks.append(msk_resized)
     arr_images = np.stack(out_images)
     arr_masks = np.stack(out_masks)
-    return torch.from_numpy(arr_images).half().to(device), torch.from_numpy(arr_masks).half().to(device)
+    return torch.from_numpy(arr_images).half(), torch.from_numpy(arr_masks).half()
 
 @spaces.GPU(duration=200)
 def inference_and_return_video(dilation_iterations, num_inference_steps, video_state):
@@ -189,7 +189,8 @@ def inference_and_return_video(dilation_iterations, num_inference_steps, video_s
     images = np.array(images)
     masks = np.array(masks)
     img_tensor, mask_tensor = preprocess_for_removal(images, masks)
-    mask_tensor = mask_tensor[:,:,:,:1]
+    img_tensor=img_tensor.to("cuda")
+    mask_tensor = mask_tensor[:,:,:,:1].to("cuda")
 
     if mask_tensor.shape[1] < mask_tensor.shape[2]:
         height = 480