Class Qwen3VLVisionTower

Index

Accessors

Methods

debugRunBlocks dispose encode encodeStages load

Accessors

cacheIdentity

get cacheIdentity(): string
Computed once, under native preparation. Exact weight content prevents cross-artifact SSD reuse even when model aliases and shapes are identical.

Returns string
- Defined in packages/inference/src/models/vision/qwen3vl.ts:64

Methods

debugRunBlocks

debugRunBlocks(
    hF32: Float32Array,
    gridH: number,
    gridW: number,
    throughMerger?: boolean,
): Float32Array
Parity-bisect helper: run ONLY the block chain (+ optional merger) on a caller-provided hidden grid — used to separate patch-embed seed ulp from block-math divergence in the stage gates.
Parameters
- hF32: Float32Array
- gridH: number
- gridW: number
- throughMerger: boolean = false
Returns Float32Array
- Defined in packages/inference/src/models/vision/qwen3vl.ts:475

dispose

dispose(): void
Returns void
- Defined in packages/inference/src/models/vision/qwen3vl.ts:562

encode

encode(
pp: Qwen3VLPreprocessed,
tap?: (name: string, arr: MlxArray) => void,
): MlxArray
Encode one preprocessed image → [imageTokens, 5120] bf16.
Parameters
- pp: Qwen3VLPreprocessed
- Optionaltap: (name: string, arr: MlxArray) => void
Returns MlxArray
- Defined in packages/inference/src/models/vision/qwen3vl.ts:502

encodeStages

encodeStages(
pp: Qwen3VLPreprocessed,
): Map<string, Float32Array<ArrayBufferLike>>
Stage-instrumented encode for parity bisection: returns named f32 copies of each pipeline stage alongside the final output.
Parameters
- pp: Qwen3VLPreprocessed
Returns Map<string, Float32Array<ArrayBufferLike>>
- Defined in packages/inference/src/models/vision/qwen3vl.ts:458

`Static`load

load(modelDir: string): Qwen3VLVisionTower
Parameters
- modelDir: string
Returns Qwen3VLVisionTower
- Defined in packages/inference/src/models/vision/qwen3vl.ts:194