Struct Q8ActsX4 

pub struct Q8ActsX4 {
    pub na: usize,
    pub n_blocks: usize,
    pub qs: Vec<i8>,
    pub d: Vec<f32>,
}

A quad of up to Q8K_ACTS_X4_NC Q8_0 activations, pre-interleaved into the layout llama.cpp’s ggml_quantize_mat_q8_0_4x8 writes into block_q8_0x4 (arch/arm/repack.cpp): every 32-element block’s qs in 8-byte runs, plus the per-block per-row scales. Consumed by the i8mm 4x8 GEMMs; prepared once per matmul, same hoist as Q8KActsX4.

Fields

na: usize

Real activations in the quad (≤ 4); rows na..4 are zero padding.

n_blocks: usize

Q8_0 blocks per activation (n_cols / 32).

qs: Vec<i8>

Interleaved quants, n_blocks * 128 long. Block b, 8-element run c, quad row a, lane k ↦ qs[b*128 + c*32 + a*8 + k] = acts[a].q[b*32 + c*8 + k].

d: Vec<f32>

Activation scales, n_blocks * 4 long: d[b*4 + a] = acts[a].d[b].