EncoderCache.empty constructor

EncoderCache.empty(
  1. int numLayers,
  2. int numKvHeads
)

Fresh cache sized for a stack of numLayers blocks each with numKvHeads KV-head slots (equal to the Q-head count for standard MHA; smaller for GQA).

Implementation

factory EncoderCache.empty(int numLayers, int numKvHeads) => EncoderCache(
  List<MHACache>.generate(
    numLayers,
    (_) => MHACache.empty(numKvHeads),
    growable: false,
  ),
);