User · Build

Building the engine

One target, one binary. The onebin build drops the engine, the serving server, JARVIS, vision, and the CLI into build/1bit. Here is how to get it to compile on your hardware, and the exact gotchas that stop people.

Prerequisites

$ sudo apt update && sudo apt install -y cmake ninja-build build-essential git

TheRock 7.15.0a

The Rock is the NPU-and-GPU toolchain. Install the nightly pip wheel that ships the gfx1151 code objects.

$ pip install --index-url https://rocm.nightlies.amd.com/whl-multi-arch/ \
    "rocm[libraries,devel,device-gfx1151]"
$ export THEROCK_PIP_ROOT="$HOME/.cache/pip/therock"

CMake auto-discovers TheRock: /opt/rocm-therock → $THEROCK_PIP_ROOT → ~/.cache/lemonade/bin/therock. Never point CMAKE_PREFIX_PATH at system ROCm (/opt/rocm). Always set the target architecture.

$ export CMAKE_HIP_ARCHITECTURES=gfx1151

Build

$ cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_HIP_ARCHITECTURES=gfx1151
$ cmake --build build --target onebin

The result is the single binary build/1bit. There is no standalone zaya_server target; packaged installs ship a zaya_server symlink, dispatched by argv[0]. Header-only deps (cpp-httplib, nlohmann_json, FTXUI) are fetched automatically via CMake FetchContent.

FastFlowLM (the NPU backend)

Discovery order, no source build for the first two:

1. TheRock dist (<therock-root>/bin/flm, ships in 7.14 / 7.15a)
2. FastFlowLM .deb (/opt/fastflowlm/bin/flm)
3. flm on PATH (/usr/bin/flm)
4. Submodule build (third_party/FastFlowLM) as a last resort

Optional GPU decode (Q4NX models)

$ cmake -B build -G Ninja -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_HIP_ARCHITECTURES=gfx1151 -DZAYA_ENABLE_GPU_DECODE=ON
$ cmake --build build --target zaya_gpu_decode    # -> build/libzaya_gpu_decode.so

Auto-detected at startup for Q4NX models. Without it, inference stays on CPU.

Gotchas