mlx-bun library API
    Preparing search index...

    What a gated DeltaNet block lends its recurrent cache for one call: the layer's recurrence weights and the block's glue between the convolution and the recurrence. The block keeps its projections, norms and activations. The cache keeps the convolution state, the recurrent state and the kernels that update them.

    interface GatedDeltaParameters {
        aLog: MlxArray;
        convWeight: MlxArray;
        dtBias: MlxArray;
        heads(convolved: MlxArray): GatedDeltaHeads;
    }
    Index
    aLog: MlxArray

    Per-value-head decay logarithm (A_log) [Hv].

    convWeight: MlxArray

    Depthwise causal convolution weight [convDim, K, 1].

    dtBias: MlxArray

    Per-value-head timestep bias [Hv].

    • The block's glue: the activation of the convolution output, the split into q, k and v heads, and their norms and scaling. Pure. convolved [B, S, convDim] is the raw convolution output (before activation) and is borrowed. The cache owns and disposes the returned heads. The cache calls this once per recurrence call; a speculative round keeps the window's k and v for the rows a rollback retains.

      Parameters

      Returns GatedDeltaHeads