The ggml-org team has released llama.cpp version b11710, a new build of its open-source software for running language models. Its listed change adds grouped-expert top-k fusion for CUDA, a graphics-card computing platform, and includes tests plus a HIP-related adjustment.
The release page provides downloads for several systems, including macOS, iOS and Linux. It does not report benchmark results or say how much the change affects speed or memory use.
Why it matters
Developers using CUDA with grouped-expert models can try the new implementation, but the release notes do not establish whether it improves performance in practice.
