The llama.cpp b10034 release addresses a critical issue where compilers for Adreno A7x devices miscompile repack kernels, leading to corrupted weights and garbage output in Mixture of Experts (MoE) models. To resolve this, the update excludes Adreno A7x devices from using these specific MoE kernels.
- Excludes Adreno A7x from using Adreno MoE kernels due to compiler miscompilation issues.
- Also excludes A6x and unknown Adreno GPUs from MoE weights repack to ensure stability.
- Disables KleidiAI support for macOS Apple Silicon builds.
- Provides binaries for macOS (Apple Silicon and Intel), Linux (CPU, Vulkan, ROCm, OpenVINO, SYCL), Android, Windows (CPU, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), and openEuler.
This change prevents model output corruption on affected Adreno hardware, ensuring reliable inference for users with these GPUs.