b9114

Build b9114 introduces a NCCL-free AllReduce implementation for tensor-parallel inference that pipelines device-to-host copy, cross-GPU communication, and reduction in a single CUDA kernel. The release renames the --allreduce flag to --reduction-provider for consistency with other multi-value flags in llama-bench. It also fixes the null log callback to ensure WARN and ERROR messages are always visible, since they indicate legitimate issues like unavailable reduction providers.

Source: llama.cpp Releases | 2026-05-12