nx_gen_rope_verify.nx

nx_gen_rope_verify.nx -- SOVEREIGN interleaved RoPE, verified vs the oracle. y[h][l][2j+r] = x[l][h][2j] * pe[l][j][r][0] + x[l][h][2j+1] * pe[l][j][r][1] pe holds a 2x2 rotation per (token l, pair j): [[cos, -sin], [sin, cos]], so r=0 gives x0*cos - x1*sin and r=1 gives x0*sin + x1*cos -- ordinary interleaved rotary embedding. Derived from rope.hpp's apply_rope (permute/reshape/repeat chain), not assumed. Usage: nx_gen_rope_verify <model> <x> <pe> <y> <head_dim> <n_heads> [rows] ⚠THE AXIS PERMUTATION IS THE WHOLE DIFFICULTY, and it is silent when wrong: input x is [head_dim, n_heads, L] -> index (l*n_heads + h)*head_dim + d output y is [head_dim, L, n_heads] -> index (h*L + l)*head_dim + d apply_rope permutes head and token axes. Reading either side in the other's order produces finite, plausible, entirely wrong numbers -- no NaN, no crash, just a silently different tensor. Both strides are therefore written out explicitly here rather than shared. pe memory index: ((l*(head_dim/2) + j)*2 + r)*2 + c -- ne = [2, 2, head_dim/2, L] Hardware __f32_* intrinsics in the hot loop, never the nx_f32_* software twins. license_tier: ORIGINAL