Class TurboQuantCodec

constructor

• new TurboQuantCodec(
      kBits: number,
      vBits: number,
      fusedDecode: boolean,
  ): TurboQuantCodec

  Parameters

  • kBits: number
  • vBits: number
  • fusedDecode: boolean

  Returns TurboQuantCodec

ReadonlyfusedDecode

ReadonlykBits

ReadonlyvBits

decode

• decode(
      t: TurboQuantTensor,
      upTo: number,
      headDim: number,
      deferV?: boolean,
  ): [MlxArray, MlxArray]

  Decode a TurboQuantTensor's active [B,H,upTo,*] window back to bf16 [B,H,upTo,headDim] (k, v). Caller owns the returned arrays. deferV leaves V in the rotated domain (decodeValuesRotated) — the caller un-rotates its attention output instead (tq.unrotateValues).

  Parameters

  • t: TurboQuantTensor
  • upTo: number
  • headDim: number
  • deferV: boolean = false

  Returns [MlxArray, MlxArray]

encode

• encode(k: MlxArray, v: MlxArray): TurboQuantTensor

  Encode k/v [B,H,L,D] into a fresh TurboQuantTensor (packed at rest).

  Parameters

  • k: MlxArray
  • v: MlxArray

  Returns TurboQuantTensor