Class Qwen3Layer

Index

Constructors

Properties

inputNorm isLinear linearAttn mlp postAttnNorm selfAttn

Methods

forward forwardAttn forwardMlp

Constructors

constructor

new Qwen3Layer(
    weights: Weights,
    config: ModelConfig,
    layerIdx: number,
    representations?: TrellisRepresentations,
): Qwen3Layer
Parameters
- weights: Weights
- config: ModelConfig
- layerIdx: number
- Optionalrepresentations: TrellisRepresentations
Returns Qwen3Layer
- Defined in packages/inference/src/models/qwen/blocks.ts:531

Properties

`Readonly`inputNorm

inputNorm: RMSNorm

`Readonly`isLinear

isLinear: boolean

`Readonly`linearAttn

linearAttn: GatedDeltaNet<QuantizedLinear> | null = null

`Readonly`mlp

`Readonly`postAttnNorm

postAttnNorm: RMSNorm

`Readonly`selfAttn

selfAttn: Qwen3Attention<QuantizedLinear> | null = null

Methods

forward

forward(
    x: MlxArray,
    cache: Cache,
    phase: ReadPhase,
    mrope: MropeForwardState | null,
    width: number,
): MlxArray
One layer over x [.., hidden] (borrowed) for phase; width is the class of x's rows, which picks the MLP's kernels.
Parameters
- x: MlxArray
- cache: Cache
- phase: ReadPhase
- mrope: MropeForwardState | null
- width: number
Returns MlxArray
- Defined in packages/inference/src/models/qwen/blocks.ts:544

forwardAttn

forwardAttn(
    x: MlxArray,
    cache: Cache,
    phase: ReadPhase,
    mrope?: MropeForwardState | null,
): MlxArray
Parameters
- x: MlxArray
- cache: Cache
- phase: ReadPhase
- mrope: MropeForwardState | null = null
Returns MlxArray
- Defined in packages/inference/src/models/qwen/blocks.ts:549

forwardMlp

forwardMlp(h: MlxArray, phase: ReadPhase, width: number): MlxArray
Parameters
- h: MlxArray
- phase: ReadPhase
- width: number
Returns MlxArray
- Defined in packages/inference/src/models/qwen/blocks.ts:557