Building from source

Building from source

You only need this to change the code. The installer builds from source by itself on older systems.

On the board

sudo apt install git cmake build-essential libssl-dev curl
git clone https://github.com/anand34577/ggml-rknpu2.git
cd ggml-rknpu2
scripts/build.sh
dist/ggml-rknpu2/llama-npu check

scripts/build.sh downloads the llama.cpp version pinned in versions.env, adds the backend, downloads Rockchip's runtime and builds, tuned for the board's CPU. It takes 10 to 40 minutes. scripts/build.sh --help lists the options, for example --portable for one build that runs on every supported chip, or --system-runtime to use /usr/lib/librknnrt.so.

Tests, on any Linux PC

No board needed. The tests replace Rockchip's runtime with a strict stand-in:

tests/run.sh      # backend correctness on all five chips
tests/e2e.sh      # a small model with and without the NPU path; the answers must match

Board tools

tools/ has two programs that measure the NPU directly. Build them on the board after scripts/build.sh; the comment at the top of each file says how.