llama.cpp Release b12373 Adds Qualcomm Hexagon SSM_SCAN Support
Llama.cpp release b12373 introduces support for the SSM_SCAN operation on Qualcomm Hexagon hardware. The update optimizes HVX loops, vectorizes softplus, and integrates HMX support to improve efficiency on Snapdragon edge devices. This release enhances local AI performance and hardware acceleration on mobile and edge devices utilizing Qualcomm Snapdragon processors. It expands the range of model architectures that can run efficiently on specialized mobile NPUs and DSPs. The pull request optimizes the DMA pipeline to match other operations, removes fallback paths in favor of direct DMA/VTCM usage, and improves the multi-device row-split solver. These changes specifically target Qualcomm's Hexagon digital signal processors and neural processing units.
## BACKGROUND
Llama.cpp is an open-source software library that enables efficient inference for large language models on local hardware using the GGML tensor library. Qualcomm Hexagon is a family of digital signal processors and neural processing units designed to deliver high performance with low power consumption on mobile platforms.