According to @_avichawla, NVIDIA’s SparDA adds a Forecast head to prefetch KV blocks, delivering 1.7x faster decode and +6.5 long-reasoning points.