nx_zimage_flashattn_verify.nx
buildroot/runtime/nx_zimage_flashattn_verify.nx
about
nx_zimage_flashattn_verify.nx -- SOVEREIGN FlashAttention (online-softmax) SDPA, verified EXACT vs oracle.
The O(n)-memory attention: instead of building the n x n score matrix, softmaxing, then weighting V,
we STREAM keys into a running accumulator, rescaling it by exp(m_old - m_new) as the running max grows.
Peak memory is O(d) per query (the accumulator + 3 scalars m,l,a), NOT O(n) -- so it scales to any n
without materializing scores. This organ proves the recurrence is NUMERICALLY IDENTICAL to standard
softmax attention (verifies vs the same oracle ao_pre as the direct SDPA). Lossless, not an approximation.
Uses a LOCAL accumulator (no mmap scratch in the loop) per the nx_cc codegen lesson. No 3rd party.
license_tier: ORIGINAL
dependencies 7 imports · 0 importers
imports: nx_syscalls.nxnx_le.nxnx_f32.nxnx_f32_div.nxnx_f32_cvt.nxnx_f32_exp.nxnx_strconv.nx
imported by: nobody (leaf or entry point)
call flow from main pre-order; caps 40 nodes / depth 6 declared; ↻ = already shown
structs
| none |
consts
| 17 | const K_MAGIC_3840: i64 = 3840 |
| 18 | const K_MAGIC_11520: i64 = 11520 |
functions
| 20 | func zfa_load(name: *u8, nl: i64, n_floats: i64) -> *u8 |
| 44 | func main() -> i64 |