Strange "Irregular nested operations" error with #[sequential inner] attribute

[mortvest]

This code produces a following error for all gpu backends:

Known compiler limitation encountered.  Sorry.
Revise your program or try a different Futhark compiler.
Cannot handle un-sliceable allocation size: (((thread; #groups=num_groups_top_18368; groupsize=group_size_18365; full []),
  [gid_flat_18369, ltid_flat_19157]), bytes_19483,)
Likely cause: irregular nested operations inside parallel constructs.

[athas]

It's a known limitation, not a bug. Do you think it's wrong about it?

[mortvest]

Why is it produced? There is no irregular nested parallelism. The same library compiles fine as a part of a bigger program, or if I remove the #[sequential_inner] attribute.