mlx-bun library API
    Preparing search index...
    • Port of BatchRotatingKVCache.make_mask for the DECODE case. Returns a [B,1,N,S] bool mask (S == min(maxSize-1, offsetScalar) + N) for the step ABOUT to be written: causal AND sliding-window AND per-row left-padding, rolled to the ring's physical layout once wrapped.

      Inputs are the PRE-write scalar ring state (idx=_idx, offsetScalar=_offset, rotated) and per-row leftPad. The leftPad adjustments here are LOCAL to the mask (mlx-lm decrements left_padding both in make_mask AND in update; the update's decrement is the persistent one, applied separately by the cache).

      Parameters

      • B: number
      • N: number
      • leftPad: number[]
      • maxSize: number
      • window: number
      • idx: number
      • offsetScalar: number
      • rotated: boolean
      • block: boolean = ...

      Returns MlxArray