While testing both eager and aot_eager modes, I observed that aot_eager does not capture exceptions as expected. I found that during graph optimization, the torch.cuda.synchronize() call is removed in ...