llama.cpp Release b11124 Fixes CUDA Top-k MoE Execution Bug
llama.cpp released patch version b11124, introducing a CUDA backend fix to ensure top-k Mixture of Experts (MoE) kernels always fire correctly during execution. The change resolves issue PR #28432 where MoE top-k routing could fail to trigger properly. This fix ensures proper routing and computation for MoE architectures, such as Mixtral or DeepSeek models, when executed on NVIDIA GPUs via CUDA. Unreliable expert routing can lead to silent failure modes or incorrect inference outputs. The patch targets CUDA execution logic specifically for top-k expert selection in MoE models. Release b11124 also includes pre-built binaries across multiple operating systems and architectures, including Linux CUDA 12/13, Windows, macOS, Android, and Snapdragon devices.
## BACKGROUND
llama.cpp is a high-performance open-source C/C++ framework for running Large Language Models locally across various hardware backends. Mixture of Experts (MoE) is a neural network architecture that relies on top-k routing to pass tokens only to a small subset of specialized sub-networks ('experts') rather than activating the entire model for every token.