Local Ai
b9112
Release b9112 of llama.cpp introduces a NCCL-free AllReduce implementation for LLAMA_SPLIT_MODE_TENSOR using a single-phase CUDA kernel, and adds an --allreduce flag to llama-bench to select between A
Release b9112 of llama.cpp introduces a NCCL-free AllReduce implementation for LLAMA_SPLIT_MODE_TENSOR using a single-phase CUDA kernel, and adds an --allreduce flag to llama-bench to select between AllReduce providers. This is useful for isolating hangs or regressions in tensor-parallel mode.
Source: llama.cpp Releases | 2026-05-11