#batched-decode

llama.cpp tries bitmaps for faster sparse KV cell lists

Per-sequence cell bitmaps cut host-side list rebuild cost in multi-sequence decode while matching split-cache perplexity.