Class AudioTower

Readonlycfg

ReadonlyembedScale

cacheIdentity

• get cacheIdentity(): string

  Immutable encoder weights and computation settings, hashed once on use.

  Returns string

dispose

• dispose(): void

  Returns void

features

• features(mel: Float32Array, frames: number, preDivide?: boolean): MlxArray

  Mel features [frames, 128] f32 (row-major, all frames valid) → language-space soft tokens [1, nSoft, textHidden] f32, PRE-DIVIDED by embedScale (forwardEmbeddings re-multiplies — vision-tower convention; the T1 golden is the UN-divided embed_audio output, so the parity test re-multiplies). nSoft = ceil(ceil(frames/2)/2)… i.e. two stride-2 convs.

  preDivide=false returns the RAW f32 embed_audio output (the T1 golden reference point). The T2 merge path needs it because the oracle divides AFTER the bf16 cast — features.astype(bf16) / embed_scale, where mlx's weak python scalar takes the array dtype, i.e. a bf16(embedScale) divisor (verified against the oracle venv). Dividing here in f32 and casting after is up to a bf16 ulp different, which the exact greedy gate does not tolerate — the builder casts first, then divides via scalarLike (same weak-scalar semantics).

  Parameters

  • mel: Float32Array
  • frames: number
  • preDivide: boolean = true

  Returns MlxArray

Staticload

• load(modelDir: string, cfg: AudioTowerConfig, embedScale: number): AudioTower

  Load the audio_tower.* + embed_audio.* tensors from the bf16 sidecar. embedScale = the language model's embed_scale (= sqrt(textHidden)); features come out pre-divided by it.

  Parameters

  • modelDir: string
  • cfg: AudioTowerConfig
  • embedScale: number

  Returns AudioTower