mlx-bun library API
    Preparing search index...

    The graph's operations and construction facts. Pooled embedding support is declared by its operation alone. Other operations accompany the promises in graphCapabilities; composition checks their pairing when it binds (declaredGraph).

    Implements

    Index
    appendKnown: MlxKnownOperations

    Known tokens: this graph declares no committed spans, so they read as a window.

    args: UniversalArgs
    cacheLayers: readonly MlxCacheLayer[]

    One cache per layer (a ring on a sliding layer with a window), sharing one mask memo. The full-attention caches' TurboQuant read is dense (denseTurboQuantKernels): the read decodes keys and values first (the softcapped read attends the decoded window).

    config: ModelConfig

    One new token per sequence: ids [sequences, 1].

    Unscaled token embeddings, for media prompts that splice embeddings.

    encodedKvAttention: boolean

    Whether every attention layer, as bound, reads encoded (affine) KV views.

    finalNorm: AnyNorm
    layers: UniversalLayer[]
    lmHead: AnyLinear | null

    Each row count names its fixed output head.

    logitsFromHidden?: MlxHeadForward
    loraState: LoraState = ...
    multiSequencePrompt: readonly MlxPromptRange[]
    multiSequencePromptChunk: readonly MlxPromptRange[]
    prefixBase: "model"
    prompt: readonly MlxPromptRange[]

    One whole chunk of a prompt; rows are sequences times the chunk size. The GPU is this graph's one unit.

    promptChunk: readonly MlxPromptRange[]

    Optional chunk bindings when a file distinguishes chunks from prompt tails.

    ropes: UniversalRope[]
    sequence: readonly MlxSequenceRange[]

    Full-sequence scoring operations, indexed by total token rows.

    sequenceLogits: readonly MlxHeadRange[]
    sharedRope: UniversalRope

    A speculative window of depth + 1 positions per sequence. This graph declares no hidden-layer taps, so capture is always null.

    weightsBytes: number