mlx-bun library API
    Preparing search index...

    Interface TrellisQuantizeOptions

    interface TrellisQuantizeOptions {
        batch?: number;
        bits?: number;
        downAxis?: DownAxis;
        interleave?: boolean;
        kMap?: { budget?: string; path: string };
        layers?: number;
        ldlq?: string;
        reuse?: string[];
        seed?: number;
        shardBytes?: number;
    }
    Index
    batch? bits? downAxis? interleave? kMap? layers? ldlq? reuse? seed? shardBytes?
    batch?: number

    Coded rows per Viterbi batch; a memory knob that never changes the codes. Default 16384.

    bits?: number

    Trellis bits per weight for every coded tensor without a k-map entry. Default 3.

    downAxis?: DownAxis

    Dim of down_proj the trellis runs along. Default out (the rotated dim).

    interleave?: boolean

    Reorder eligible k=3 axis-0 codes into the two-block scatter layout.

    kMap?: { budget?: string; path: string }

    Per-tensor bits (and affine overrides) from an allocation file's budgets[budget].

    layers?: number

    Trellis-code only the first N decoder layers (the rest ride the affine base tier).

    ldlq?: string

    Directory of per-layer Hessian factors; enables BlockLDLQ.

    reuse?: string[]

    Packed artifacts to copy unchanged-geometry tensors from, first match wins.

    seed?: number

    Rotation seed (the MTP companion must use its trunk's). Default 42.

    shardBytes?: number

    Output shard size in bytes. Default 2 GiB.

    Settings
    Member Visibility

    On This Page
    Properties