b9101

Build b9101 of llama.cpp includes improvements to the AllReduce functionality for tensor-parallel GPU inference, introducing a NCCL-free implementation using CUDA kernels for cross-GPU communication. It also renames the llama-bench reduction provider flag from --allreduce to --reduction-provider/-rp for consistency with other multi-value options. The build improves logging by ensuring WARN and ERROR messages display in non-verbose mode to surface legitimate issues like unavailable reduction providers.

Source: llama.cpp Releases | 2026-05-10