pub fn sparse_attn(
q: &[f32],
kv: &[f32],
sink: &[f32],
idx: &[i32],
s: usize,
h: usize,
d: usize,
topk: usize,
scale: f32,
) -> Vec<f32>Expand description
2026-09-25: sparse_attn as in ds41_ref_shims.py: per query and head,
softmax over the gathered rows plus the sink (denominator only), all in
f32; output bf16. q: [s][h][d], kv: [rows][d], idx: [s][topk]
with -1 skipped.