The new llama.cpp release addresses critical miscompilation issues with MoE repack kernels on Adreno A7x and A6x GPUs, which previously corrupted weights and caused garbage output for MoE models. This update also excludes unknown Adreno devices from using MoE weights repack to prevent similar issues. The release provides updated build targets across macOS, iOS, Linux, Android, and Windows, supporting various hardware backends including Apple Silicon, Vulkan, ROCm, OpenVINO, SYCL, CUDA, and HIP.
Corrected MoE inference on Adreno GPUs improves reliability for mobile and edge AI deployments, expanding viable hardware for LLM inference.