Class QuantizedSwitchLinear

Port of switch_layers.QuantizedSwitchLinear (gather_qmm over stacked expert weights; rhs_indices selects the expert per row).

Index

Constructors

Properties

biases scales spec w

Methods

Constructors

constructor

new QuantizedSwitchLinear(
    w: MlxArray,
    scales: MlxArray,
    biases: MlxArray | null,
    spec: QuantSpec,
): QuantizedSwitchLinear
Parameters
- w: MlxArray
- scales: MlxArray
- biases: MlxArray | null
- spec: QuantSpec
Returns QuantizedSwitchLinear
- Defined in packages/inference/src/layers/quantized-switch-linear.ts:10

Properties

`Readonly`biases

biases: MlxArray | null

`Readonly`scales

scales: MlxArray

`Readonly`spec

spec: QuantSpec

`Readonly`w

Methods

forward

forward(x: MlxArray, indices: MlxArray, sortedIndices: boolean): MlxArray
Parameters
- x: MlxArray
- indices: MlxArray
- sortedIndices: boolean
Returns MlxArray
- Defined in packages/inference/src/layers/quantized-switch-linear.ts:34

`Static`load

load(weights: Weights, path: string, config: ModelConfig): QuantizedSwitchLinear
Parameters
- weights: Weights
- path: string
- config: ModelConfig
Returns QuantizedSwitchLinear
- Defined in packages/inference/src/layers/quantized-switch-linear.ts:17