Struct W4a4Proj 

pub struct W4a4Proj(/* private fields */);

2026-09-30: The W4A4 kernels of one backend (see the module header).

Implementations

impl W4a4Proj

pub fn for_launch( gpu: &dyn GpuBackend, weight: &QuantizedWeight, m: u32, n: u32, k: u32, ) -> Option<Self>

2026-09-30: The kernels for an m-row [n, k] projection over weight, when the tier admits it there (nvfp4_proj_small_m then runs W4A4); None when it runs W4A16.

pub fn prepared(gpu: &dyn GpuBackend) -> Option<Self>

2026-09-30: The kernels once prepare found them, for the quantize launch alone.

pub fn quant_kernel(&self) -> KernelHandle

2026-09-30: w4a4_quant_rows.

pub fn mx_kernel(&self, m: u32, n: u32, k: u32) -> KernelHandle

2026-09-30: The MX entry an m-row [n, k] launch takes.

pub fn quantize( &self, gpu: &dyn GpuBackend, input: DevicePtr, act: Nvfp4ActBuf, m: u32, k: u32, stream: u64, ) -> Result<()>

2026-09-30: Quantize m rows of the BF16 input ([m, k]) into act.

pub fn gemv( &self, gpu: &dyn GpuBackend, act: Nvfp4ActBuf, weight: &QuantizedWeight, output: DevicePtr, m: u32, n: u32, k: u32, stream: u64, ) -> Result<()>

2026-09-30: output[m, n] (BF16) = the W4A4 product of act and weight.

Trait Implementations

impl Clone for W4a4Proj

fn clone(&self) -> W4a4Proj

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Copy for W4a4Proj