Class AneTrellisMlpSplit

Readonlychannels

Readonlydown

Readonlygate

Readonlygpu

Readonlyprogram

Readonlyrest

Readonlyrows

Readonlyup

[dispose]

• "[dispose]"(): void

  Returns void

dispose

• dispose(): void

  Releases this layer's program and frees its GPU rows; a shared program or buffers outlive it while another layer holds them.

  Returns void

forward

• forward(hidden: MlxArray, swiglu: Swiglu): MlxArray

  hidden [.., M, D], M <= rows → MLP(hidden) [.., M, D], unevaluated (mlpSplit).

  Parameters

  • hidden: MlxArray
  • swiglu: Swiglu

  Returns MlxArray

Staticbuild

• build(
      gate: TrellisLinear,
      up: TrellisLinear,
      down: TrellisLinear,
      fraction: number,
      rows: number,
      share?: AneMlpSplit,
      slices?: number,
      gpuRows?: AneGpuMlpRows,
  ): AneTrellisMlpSplit

  Loads aneMlpSplitProgram(D, R, fraction, rows, slices) for pieces of up to rows rows now, from the compile cache the install step filled (a program not compiled on this Mac throws AneBridgeError). share: a split that never runs at the same time as this one; this split holds its program when the two are identical, otherwise loads one on its buffers (each at least as large).

  Parameters

  • gate: TrellisLinear
  • up: TrellisLinear
  • down: TrellisLinear
  • fraction: number
  • rows: number
  • Optionalshare: AneMlpSplit
  • Optionalslices: number
  • gpuRows: AneGpuMlpRows = legacyGpuRows

  Returns AneTrellisMlpSplit