You can't as far as I'm aware unless you control the entire batch during inference, or don't use batching which would require you to run your own inference.