llama.cpp

Tensor-parallel: Fix delayed AllReduce on Gemma-4 MoE

#22129

Merged