Interface FlashCceHead

The quantized head pieces the kernel needs (packed weight + per-group affine scale/bias + bits/group_size). Mirrors HeadQuant in loss.ts.

interface FlashCceHead {
    biases: MlxArray;
    bits: number;
    groupSize: number;
    scales: MlxArray;
    softcap: number | null;
    w: MlxArray;
}

Index

Properties

biases bits groupSize scales softcap w

Properties

biases

biases: MlxArray

bits

bits: number

groupSize

groupSize: number

scales

scales: MlxArray

softcap

softcap: number | null

w