Struct W8a8Ctx 

pub struct W8a8Ctx {
    pub kernels: W8a8Kernels,
    pub scratch: W8a8Scratch,
}

2026-09-28: The kernels and the quantized-activation scratch one model’s W8A8 projections share.

Fields

kernels: W8a8Kernels

scratch: W8a8Scratch

Implementations

impl W8a8Ctx

pub fn new(gpu: &dyn GpuBackend, max_k: u32) -> Result<Self>

2026-09-28: Look the kernels up and allocate a scratch for a max_k-wide activation of ops::W8A8_MAX_ROWS rows.

pub fn available(&self, w: &W8a8Weight, rows: usize) -> bool

2026-09-28: Whether w can run W8A8 at rows rows.

pub fn proj( &self, gpu: &dyn GpuBackend, w: &W8a8Weight, x: DevicePtr, ldx: u32, rows: usize, out: DevicePtr, ldc: u32, stream: u64, ) -> Result<bool>

2026-09-28: out[rows, ldc] = W8A8 of x[rows, ldx] (BF16) and w; Ok(false), launching nothing, when not Self::available.

pub fn silu_proj( &self, gpu: &dyn GpuBackend, w: &W8a8Weight, gate: DevicePtr, up: DevicePtr, ld: u32, rows: usize, out: DevicePtr, ldc: u32, stream: u64, ) -> Result<bool>

2026-09-28: The down projection of a SiLU FFN: out = W8A8 of bf16(silu(gate) * up) (gate and up [rows, ld] BF16) and w, with the SiLU product quantized in the same launch; Ok(false) when not available.

Trait Implementations

impl Clone for W8a8Ctx

fn clone(&self) -> W8a8Ctx

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for W8a8Ctx

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

Formats the value using the given formatter.

impl Copy for W8a8Ctx