ggerganov/llama.cpp Released b12361 with Code Reorganization
Release b12361 of llama.cpp introduces minor internal code reorganization by moving the MoE-cache code out of the llama-graph component. This update continues the ongoing maintenance and architectural cleanup of the codebase. While this is a routine point release without major new user-facing features, internal code refactoring helps maintain project modularity and long-term maintainability. Developers and power users relying on Mixture of Experts (MoE) models benefit from cleaner underlying graph structures. The primary change in pull request #32157 involves separating the MoE cache logic from the main computation graph implementation. The release package includes prebuilt binaries for various operating systems and hardware backends, including macOS, Linux, and Windows.
## BACKGROUND
llama.cpp is a widely used open-source software library designed to perform local inference on large language models in C and C++. Mixture of Experts (MoE) models utilize a sparse architecture where only a subset of neural network parameters or 'experts' are activated per token, improving efficiency.