llama.cpp

5503b04b - meta: clear inactive AllReduce shards with FILL, not SCALE (#29793)