Class LlamaUnfusedMLP

The MLP with the swiglu as separate silu and multiply kernels: the explicit layer form, available for callers writing their own graph.

Index

Constructors

Properties

down gate kind up

Methods

Constructors

constructor

new LlamaUnfusedMLP(
    weights: Weights,
    config: ModelConfig,
    prefix: string,
): LlamaUnfusedMLP
Parameters
- weights: Weights
- config: ModelConfig
- prefix: string
Returns LlamaUnfusedMLP
- Defined in packages/inference/src/models/minicpm5/model.ts:181

Properties

`Readonly`down

down: QuantizedLinear

`Readonly`gate

gate: QuantizedLinear

`Readonly`kind

kind: "unfused"

`Readonly`up

up: QuantizedLinear

Methods

forward

forward(x: MlxArray, residual?: MlxArray): MlxArray
MLP: returns mlp(x); the caller adds the residual when given.
Parameters
- x: MlxArray
- Optionalresidual: MlxArray
Returns MlxArray
- Defined in packages/inference/src/models/minicpm5/model.ts:188