llama.cpp b11055 Adds GEGLU_QUICK Support for Qualcomm Hexagon Backend
Open-source LLM inference engine llama.cpp has released version b11055, featuring pull request #29114. This update adds support for the GEGLU_QUICK activation function to the Qualcomm Hexagon backend. Expanding activation function support on Qualcomm Hexagon hardware allows models using GEGLU variant architectures to run efficiently on Snapdragon-powered mobile and edge devices. This improves on-device LLM inference capabilities across a broader range of neural network architectures. The release specifically integrates the `GEGLU_QUICK` variant, an optimized approximation of the Gaussian Error Gated Linear Unit activation, into the Hexagon hardware accelerator backend. Binaries for release b11055 are provided across multiple operating systems and platforms, including Android, Linux, macOS, and Windows.
## BACKGROUND
Qualcomm Hexagon is a dedicated digital signal processor (DSP) architecture built into Snapdragon processors, optimized for low-power mobile multimedia and AI workloads. GEGLU (Gaussian Error Gated Linear Unit) is a variant of the Gated Linear Unit activation function that uses GELU, widely adopted in modern Transformer architectures to enhance model capacity and training effectiveness.