Struct MlaInputs 

pub struct MlaInputs<'a> {
    pub q_nope: &'a [f32],
    pub q_pe: &'a [f32],
    pub c_kv: &'a [f32],
    pub k_pe: &'a [f32],
    pub w_uk: &'a [f32],
    pub w_uv: &'a [f32],
    pub t_q: usize,
    pub t_kv: usize,
}

Inputs shared by both forms. Rope is already applied to q_pe/k_pe (it happens upstream of the attention core and is identical in both forms). c_kv is already RMS-normed.

Layouts (row-major): q_nope: [t_q][n_head][d_nope] q_pe: [t_q][n_head][d_rope] c_kv: [t_kv][kv_rank] — the latent KV cache (one row per token, all heads) k_pe: [t_kv][d_rope] — decoupled rope key (one per token, all heads) w_uk: [n_head][d_nope][kv_rank] — k_nope_h = w_uk[h] · c_kv w_uv: [n_head][d_v][kv_rank] — v_h = w_uv[h] · c_kv

The queries occupy the LAST t_q positions of the cache (decode/prefill convention: their own rows are already appended). Causal: query i attends to cache rows 0 ..= (t_kv - t_q + i).

Fields

q_nope: &'a [f32]

q_pe: &'a [f32]

c_kv: &'a [f32]

k_pe: &'a [f32]

w_uk: &'a [f32]

w_uv: &'a [f32]

t_q: usize

t_kv: usize