llama.cpp Release b10852 Adds ReLU and Leaky ReLU Support for Hexagon Backend
The llama.cpp project released version b10852, introducing support for RELU and LEAKY_RELU activation operations within its Qualcomm Hexagon backend. The automated release also updates pre-built binaries across supported operating systems including macOS, Linux, Windows, and Android. Expanding operator coverage for the Hexagon backend allows a broader range of neural network components to execute natively on Qualcomm hardware acceleration units. This incrementally improves on-device AI inference capabilities and efficiency for Snapdragon-powered mobile and edge devices. This minor release incorporates PR #28585 to add both standard ReLU and Leaky ReLU operation implementations for Hexagon DSPs. Pre-built release artifacts are provided for various target platforms, though macOS ARM64 with KleidiAI and openEuler builds remain disabled in this tag.
## BACKGROUND
llama.cpp is a popular open-source C/C++ framework enabling efficient inference of Large Language Models across diverse hardware, from CPUs to specialized hardware backends. Qualcomm Hexagon is a Digital Signal Processor (DSP) and NPU architecture integrated into Qualcomm Snapdragon processors designed to accelerate AI workloads. ReLU and Leaky ReLU are core activation functions in deep learning models, where Leaky ReLU assigns a small non-zero slope to negative inputs to prevent neurons from becoming inactive.