Class ReduceGateUpRows

2..4 activation rows sharing each decoded weight; gate and up at the same width.

Index

Constructors

Properties

Methods

Constructors

constructor

new ReduceGateUpRows(gate: TrellisLinear, up: TrellisLinear): ReduceGateUpRows
Parameters
- gate: TrellisLinear
- up: TrellisLinear
Returns ReduceGateUpRows
- Defined in packages/inference/src/layers/trellis-reduce-gate-up.ts:42

Properties

`Readonly`gate

gate: TrellisLinear

`Readonly`up

up: TrellisLinear

Methods

forward

forward(x: MlxArray): MlxArray
x [.., in] (borrowed, 2..4 rows) to an owned [.., rows].
Parameters
- x: MlxArray
Returns MlxArray
- Defined in packages/inference/src/layers/trellis-reduce-gate-up.ts:48