pub fn mla_masked_attention(
q: &[f32],
k: &[f32],
v: &[f32],
mask: &[u8],
dims: DsaDims,
q_rows: usize,
kv_len: usize,
) -> Vec<f32>Expand description
2026-09-25: NoPE MLA over a per-query selected key set.
q: [q_rows, heads, qk_head_dim], k: [kv_len, heads, qk_head_dim],
v: [kv_len, heads, v_head_dim], mask: [q_rows, kv_len].
The scale is qk_head_dim^-0.5 over the full head dim; it equals qk_nope_head_dim^-0.5 only
while the rope part is zero-width.