ggerganov/llama.cpp released build b11912 with MoE and Laguna support
llama.cpp release b11912 introduces support for Mixture of Experts (MoE) amax rescaling and declares Laguna SM tensor support. These updates improve how specific large language model architectures handle tensor operations and scaling. This update enhances the framework's capability to efficiently run advanced Mixture of Experts models like Poolside's Laguna architecture on consumer and enterprise hardware. Such improvements ensure better compatibility and performance for newer, complex open-weight models. The release includes pre-compiled binaries across multiple platforms—including macOS, Linux, Windows, and Android—supporting diverse backends such as CPU, CUDA, Vulkan, and Snapdragon NPUs. The MoE amax rescaling implementation was assisted by Devin (#31147).
## BACKGROUND
Mixture of Experts (MoE) is a neural network architecture where different subnetworks (experts) handle different subsets of data, improving efficiency over a single dense layer. llama.cpp is a popular open-source framework designed to run large language models locally with minimal resource overhead across various hardware backends.