llama.cpp

8b0e0db6 - TP: fix granularity for Qwen 3.5/3.6 + 3 GPUs (#23843)