Class Glm52MtpGraph

constructor

• new Glm52MtpGraph(model: Glm52MtpHost, layer: Glm52MtpLayer): Glm52MtpGraph

  Parameters

  • model: Glm52MtpHost
  • layer: Glm52MtpLayer

  Returns Glm52MtpGraph

Readonlylayer

Readonlymodel

cache

• get cache(): { kvLoraRank: number; maxTokens: number; ropeHeadDim: number }

  Geometry of the compressed cache the head appends to.

  Returns { kvLoraRank: number; maxTokens: number; ropeHeadDim: number }

hiddenSize

• get hiddenSize(): number

  Returns number

decode

• decode(
      ids: MlxArray,
      hidden: MlxArray,
      cache: LatentAttentionCache,
  ): Promise<MlxArray>

  One token per sequence (a draft step): the MTP layer reads its compressed state absorbed.

  Parameters

  • ids: MlxArray
  • hidden: MlxArray
  • cache: LatentAttentionCache

  Returns Promise<MlxArray>

project

• project(hidden: MlxArray): MlxArray

  Logits [B, positions, V] of the head's output.

  Parameters

  • hidden: MlxArray

  Returns MlxArray

window

• window(
      ids: MlxArray,
      hidden: MlxArray,
      cache: LatentAttentionCache,
  ): Promise<MlxArray>

  Two or more tokens per sequence (the accepted window a commit replays): reconstructed.

  Parameters

  • ids: MlxArray
  • hidden: MlxArray
  • cache: LatentAttentionCache

  Returns Promise<MlxArray>