Struct DenseWeight 

pub struct DenseWeight {
    pub weight: DevicePtr,
}

BF16 dense weight (no quantization).

Fields

weight: DevicePtr

Implementations

impl DenseWeight

pub fn quantize_to_fp8( &self, gpu: &dyn GpuBackend, quantize_kernel: KernelHandle, n: usize, k: usize, stream: u64, ) -> Result<Fp8DenseWeight>

Quantize a BF16 weight [N, K] to FP8 E4M3 [N, K] with per-row f32 scales. Allocates the FP8 buffer + row_scale buffer on the GPU, runs the quantize_bf16_to_fp8 kernel, and returns the resulting Fp8DenseWeight.

Called once at model load time. Caller is responsible for any stream synchronization needed before the returned weight is consumed by fp8_gemm_n128 or related kernels.

Phase G (DFlash drafter FP8 weights). Mirrors QuantizedWeight::predequant_to_fp8 for the BF16 source path.

Trait Implementations

impl Clone for DenseWeight

fn clone(&self) -> DenseWeight

Returns a duplicate of the value.

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

Performs copy-assignment from source.

impl Debug for DenseWeight

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

Formats the value using the given formatter.

impl From<DenseWeight> for QuantWeight

fn from(w: DenseWeight) -> Self

Converts to this type from the input type.

impl Copy for DenseWeight