llama.cpp

TP: fix 0-sized tensor slices, AllReduce fallback

#21808

Merged