Class VisionTower

ReadonlyembedScale

ReadonlyrmsNormEps

cacheIdentity

• get cacheIdentity(): string

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

  Returns string

dispose

• dispose(): void

  Returns void

features

• features(img: PreprocessedImage): MlxArray

  One image's preprocessed patches → language-space features [softTokens, hidden] (pre-divided by embed_scale).

  Parameters

  • img: PreprocessedImage

  Returns MlxArray

preprocess

• preprocess(bytes: Uint8Array): Promise<PreprocessedImage>

  Parameters

  • bytes: Uint8Array

  Returns Promise<PreprocessedImage>

Staticload

• load(modelDir: string, embedScale: number, rmsNormEps?: number): VisionTower

  Parameters

  • modelDir: string
  • embedScale: number
  • rmsNormEps: number = 1e-6

  Returns VisionTower