One forward over a prefix-shared concat instead of two [prompt; response] forwards: block-sparse mask plus block-wise rope, bit-exact with the pair.
Post-final-norm hidden [1, T, hidden] of the concat ids under layout. Borrows ids; the caller owns the result.
ids
layout
One forward over a prefix-shared concat instead of two [prompt; response] forwards: block-sparse mask plus block-wise rope, bit-exact with the pair.