Running llama.cpp on Snapdragon Hexagon NPU seems promising

https://github.com/ggml-org/llama.cpp/blob/master/docs/backend/snapdragon/README.md I have an Oneplus 12 with Snapdragon 8 Gen 3. I followed the above README to cross-compile llama.cpp on Ubuntu and then copy to the Termux directory on the phone. It seems like llama.cpp's Hexagon backend is highly supported by Qualcomm with many PRs made by Qualcomm employees. I am getting 8t/s pp and 4.5t/s tg with gemma-3-12b-it-qat-Q4_0 and 20t/s pp 12.5t/s tg with gemma-3-4b-it-qat-Q4_0. Speed is about t