This operation requires a considerable amount of lines and has many pitfalls (e.g. `"cuda:2"`). It needs to be replicated for every supported engine.