Struct DecodeW8a8Plan 

pub struct DecodeW8a8Plan {
    pub rows: usize,
    pub n: u32,
    pub k: u32,
    pub ldc: u32,
    pub out_capacity_bytes: usize,
}

Everything the decode W8A8 cuBLASLt arm needs, as a PURE function of shape, format, lever and handles — so the CPU tests pin every clause without a GPU and without touching the process environment.

Clauses, each load-bearing:

• family_armed — the caller’s slice of super::CublasScope (cublas.ssm / cublas.attn). Arming the dense FFN must not arm these; that separation is the whole point of the scoped lever (#917’s 10.3 GiB).
• !disabled — the ATLAS_NO_W8A8_DECODE_PROJ kill switch.
• DECODE_W8A8_ROWS.contains(&rows) — the 5..=16 band, on the PADDED n.
• Fp8BlockScaled — cuBLASLt is told the weight scales are a BLK128x128 grid; a per-row row_scale has a different shape and reads as garbage.
• n % 128 == 0, k % 128 == 0 — that grid is [N/128, K/128] and the activation quantizer emits one scale per 128-wide K group.
• blk128x128_stride_ok(k) (i.e. k % 512 == 0) — cuBLAS requires the weight-scale column stride K/128 to be a multiple of 4.
• ldc >= n — a leading dimension shorter than the column is rejected by the library; for a contiguous output the caller passes ldc = n.
• OUTPUT ROOM for the full padded write extent (see strided_out_extent_elems).
• the quantizer kernel, and the VEC128 scale-layout adapter (kernel AND scratch AND its capacity, and the FP8/scale scratch capacities). cuBLASLt reads the activation scales token-contiguous; handing over the quantizer’s [M, K/128] order is fast and WRONG (H100 2026-09-11: rel_rms 7.7e-2 / ~33 000 BF16 ULP on identical FP8 bytes). Falling back to the batch16 GEMV is the only safe answer when any of it is missing.

Fields

rows: usize

Padded decode rows — the ctx n the CUDA-graph cache is keyed by.

n: u32

Output width of this projection.

k: u32

Contract width of this projection.

ldc: u32

Output row pitch in BF16 ELEMENTS (n for a contiguous output).

out_capacity_bytes: usize

Allocated size of the output buffer, in bytes.

Implementations

impl DecodeW8a8Plan

pub fn contiguous( rows: usize, n: u32, k: u32, out_capacity_bytes: usize, ) -> Self

A contiguous [rows, n] output (SSM in_proj_qkvz, SSM out_proj, attention o_proj).

pub fn strided( rows: usize, n: u32, k: u32, ldc: u32, out_capacity_bytes: usize, ) -> Self

A strided output: rows ldc BF16 elements apart (attention Q/K/V into the [n, per_seq_qkv] multi-seq QKV buffer).

pub fn m_pad(&self) -> u32

The padded M cuBLASLt is actually handed.

pub fn write_extent_bytes(&self) -> usize

Bytes of out this projection may touch, phantom rows included.

Trait Implementations

impl Clone for DecodeW8a8Plan

fn clone(&self) -> DecodeW8a8Plan

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Debug for DecodeW8a8Plan

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Copy for DecodeW8a8Plan