Under inline / fast-path PyTask execution, cuPyNumeric's LAPACK-backed tasks receive row-major
(C-order) inputs, but LAPACK requires column-major (Fortran) layout. Inline execution bypasses
the mapper, so the Fortran layout that cuPyNumeric normally requests for LAPACK tasks
is never applied; the store is instead created as hardcoded C-order.
This affects solve, qr, svd and any other cuPyNumeric task that relies on the
mapper honoring its column-major layout request.
Under inline / fast-path PyTask execution, cuPyNumeric's LAPACK-backed tasks receive row-major
(C-order) inputs, but LAPACK requires column-major (Fortran) layout. Inline execution bypasses
the mapper, so the Fortran layout that cuPyNumeric normally requests for LAPACK tasks
is never applied; the store is instead created as hardcoded C-order.
This affects
solve,qr,svdand any other cuPyNumeric task that relies on themapper honoring its column-major layout request.