Class Trellis

constructor

• new Trellis(cfg: TrellisCfg): Trellis

  Parameters

  • cfg: TrellisCfg

  Returns Trellis

Readonlycfg

ReadonlyG

ReadonlyJ

ReadonlylutRms

ReadonlynStates

decodeStates

• decodeStates(idx: MlxArray): MlxArray

  States → decoded values ([B,T] f32), the exact map encodeDecode uses.

  Parameters

  • idx: MlxArray

  Returns MlxArray

dispose

• dispose(): void

  Returns void

encodeDecode

• encodeDecode(X: MlxArray): MlxArray

  Trellis-encode + decode [B,T] rows (already variance-matched to the code). Returns [B,T] f32 reconstruction.

  Parameters

  • X: MlxArray

  Returns MlxArray

encodeDecodeChunked

• encodeDecodeChunked(X: MlxArray, maxB: number): MlxArray

  encodeDecode over [B,T] in chunks of at most maxB rows. The caller has already variance-matched X to the code (BlockLDLQ path).

  Parameters

  • X: MlxArray
  • maxB: number

  Returns MlxArray

encodeStates

• encodeStates(X: MlxArray): MlxArray

  Trellis-encode [B,T] rows (already variance-matched) and return the decoded STATE at every t as [B,T] int32 — the k low bits of state t are the emitted symbol, and (bitshift trellis) state t is exactly the last L/k symbols, which is what makes random-access decode possible.

  Parameters

  • X: MlxArray

  Returns MlxArray

fakeQuantRows

• fakeQuantRows(
      W: MlxArray,
      batchBlocks: number,
      onProgress?: (done: number, total: number) => void,
  ): MlxArray

  Fake-quant a 2-D tensor whose trellis axis is the LAST axis. Per-row (per rotated vector) fp16 scale, QTIP's Wscale = RMS(W)/RMS(lut) applied per row instead of per tensor (deviation: our incoherence processing is one-sided, so rows do not share a norm).

  Parameters

  • W: MlxArray
  • batchBlocks: number
  • OptionalonProgress: (done: number, total: number) => void

  Returns MlxArray

fakeQuantRowsPacked

• fakeQuantRowsPacked(
      W: MlxArray,
      batchBlocks: number,
      onProgress?: (done: number, total: number) => void,
  ): { codes: MlxArray; rec: MlxArray; scales: MlxArray }

  fakeQuantRows that ALSO returns the packed bit-stream and the fp16 row scales — the on-disk trellis format (see packStates). rec is the same f32 reconstruction fakeQuantRows returns, so the fake-quant and packed artifacts are built from one Viterbi pass.

  Parameters

  • W: MlxArray
  • batchBlocks: number
  • OptionalonProgress: (done: number, total: number) => void

  Returns { codes: MlxArray; rec: MlxArray; scales: MlxArray }