mlx-bun library API
    Preparing search index...

    Type Alias Vision

    Prepared embeddings for vision/audio prompts. Qualified models consume this through the media prompt input the graph binds (bindMediaInput). Token-only prefix caching remains disabled for media.

    type Vision = {
        embeddings: MlxArray;
        imageMask?: MlxArray;
        mrope?: MropeRequestState;
        multimodalMask?: MlxArray;
        prefixIdentity?: string;
    }
    Index
    embeddings imageMask? mrope? multimodalMask? prefixIdentity?
    embeddings: MlxArray
    imageMask?: MlxArray

    bool [L] image-token mask for the bidirectional attention overlay. Absent when the prompt carries ANY audio — audio(-containing) prompts run fully causal (02d723a:docs/design/generic-model-support.md §6.6).

    mrope?: MropeRequestState

    Qwen3.5/3.8 vision: the request's mRoPE positions + decode delta, owned by the media prompt input the graph binds.

    multimodalMask?: MlxArray

    bool [L] union multimodal soft-token mask (image | audio) for per-layer-input id zeroing. Absent on the legacy vision-only shape, where zeroing falls back to imageMask.

    prefixIdentity?: string

    Producer identity includes all media and preceding rendered tokens.

    Settings
    Member Visibility

    On This Page
    Properties