Interface HeadQuant

The quantized LM head and final-logit softcap the fused linear-CE heads read. The head is not a LoRA target, so the base weights are authoritative.

interface HeadQuant {
    biases: MlxArray | null;
    scales: MlxArray;
    softcap: number | null;
    spec: QuantSpec;
    w: MlxArray;
}

Index

Properties

biases scales softcap spec w

Properties

biases

biases: MlxArray | null

scales

scales: MlxArray

softcap

softcap: number | null

spec

spec: QuantSpec

w