Installing whisper.cpp libraries for bucky

bucky loads whisper.cpp at runtime via purego and jupiterrider/ffi — there is no CGo in this repository. That means you need a prebuilt shared library on disk before any FFI call will succeed.

Set BUCKY_LIB (or pass -lib <path>) to the directory that contains the shared library. The expected filenames are:

OS	Filename
linux / freebsd	libwhisper.so
darwin	libwhisper.dylib
windows	whisper.dll (plus ggml*.dll siblings)

macOS (arm64 / amd64)

make build
./bucky install -lib ./lib

This downloads the official whisper-vX.Y.Z-xcframework.zip from the whisper.cpp GitHub releases and extracts the macos-arm64_x86_64 slice of the universal Mach-O dylib into ./lib/libwhisper.dylib. Metal acceleration is included.

Windows (amd64)

make build
.\bucky.exe install -lib .\lib

This downloads whisper-bin-x64.zip and extracts the Release/*.dll files (whisper.dll, ggml.dll, ggml-base.dll, ggml-cpu.dll).

For CUDA builds use -p cuda; that downloads whisper-cublas-12.4.0-bin-x64.zip instead. You must already have the matching CUDA 12.4 runtime installed on the host.

Windows ABI verification. pkg/whisper.WhisperFullParams is sized assuming LLP64 with 4-byte int and 8-byte size_t/pointer — exactly what MSVC produces on Windows amd64. The Windows GitHub Actions job runs the full FFI smoke on every push: bucky install, bucky model get tiny, go test -count=1 ./... (which exercises TestWhisperFullParamsSize, TestVadParamsSize, TestVadContextParamsSize, and TestFullWithState against the real whisper.dll), and examples/hello samples/jfk.wav. Watch the badge in README.md for regressions; if you see unsafe.Sizeof(WhisperFullParams) = N, want 304 with N != 304 the _padN fields in pkg/whisper/params.go need adjustment for the Windows ABI.

Linux (amd64 / arm64)

make build
./bucky install -lib ./lib

Linux libraries are produced by the ardanlabs/bucky-builder companion repo (whisper.cpp upstream publishes no Linux release artifact at all). The builder re-runs hourly against new whisper.cpp tags and publishes six artifacts per release:

Backend	amd64	arm64
CPU	whisper-vX.Y.Z-bin-ubuntu-cpu-x64.tar.gz	whisper-vX.Y.Z-bin-ubuntu-cpu-arm64.tar.gz
CUDA 12.9	whisper-vX.Y.Z-bin-ubuntu-cuda-x64.tar.gz	whisper-vX.Y.Z-bin-ubuntu-cuda-arm64.tar.gz
Vulkan	whisper-vX.Y.Z-bin-ubuntu-vulkan-x64.tar.gz	whisper-vX.Y.Z-bin-ubuntu-vulkan-arm64.tar.gz

bucky install auto-detects CUDA via nvidia-smi and downloads the matching artifact. Pass -p vulkan to opt into the Vulkan build, or -p cpu to force the CPU bundle. CUDA arm64 targets Jetson Orin (sm_87); CUDA amd64 targets sm_86 + sm_89 (consumer Ampere / Ada GPUs).

The tarball unpacks libwhisper.so, libggml.so, libggml-base.so, libggml-cpu.so, and (for cuda / vulkan variants) libggml-cuda.so / libggml-vulkan.so into lib/. RPATH is $ORIGIN, so the libraries are self-contained regardless of where you point BUCKY_LIB.

If you'd rather build whisper.cpp yourself:

git clone https://github.com/ggml-org/whisper.cpp.git
cd whisper.cpp
git checkout v1.8.6
cmake -B build -DBUILD_SHARED_LIBS=ON
cmake --build build --config Release -j$(nproc)
mkdir -p ../bucky/lib
cp build/src/libwhisper.so ../bucky/lib/
cp build/ggml/src/libggml*.so ../bucky/lib/