I do (I think) understand what a consume operation is when it deals with atomic operation.
However, since there is no data directly linked to a fence, I wonder if there is case where it is useful to use std::memory_order_consume with an std::atomic_thread_fence ?