Class QwenFoldContext

constructor

• new QwenFoldContext(
      weights: Pick<Weights, "tensor">,
      hiddenSize: number,
      seed: number,
      gammaNames: string[],
      eager?: boolean,
  ): QwenFoldContext

  Parameters

  • weights: Pick<Weights, "tensor">
  • hiddenSize: number
  • seed: number
  • gammaNames: string[]
  • eager: boolean = true

  Returns QwenFoldContext

ReadonlyhiddenSize

apply

• apply(op: FoldOp, src: MlxArray): MlxArray

  Apply op to src, returning an EVALUATED bf16 array the caller owns. Never disposes src (the source shard map owns it).

  Parameters

  • op: FoldOp
  • src: MlxArray

  Returns MlxArray

dispose

• dispose(): void

  Returns void

gain

• gain(gamma: string): Float32Array

  The preloaded γ named gamma, read back: the gain every fold op naming it scales by.

  Parameters

  • gamma: string

  Returns Float32Array