whispercpp-vulkan-runtime

Windows x64 builds of whisper.cpp with Vulkan GPU acceleration, available as a standalone ZIP or a pip-installable wheel.

Requirements

• Windows x64
• CPU with AVX2, FMA, and F16C support
• Vulkan-capable GPU and graphics driver
• Microsoft Visual C++ v14 x64 Redistributable
• Python 3.10 or later for the pip package only

Models are downloaded separately. Running the prebuilt runtime does not require a compiler or Vulkan SDK.

Installation

Python package

python -m pip install https://github.com/CryptoKey98/whispercpp-vulkan-runtime/releases/download/v1.8.5.1/whispercpp_vulkan_runtime-1.8.5.1-py3-none-win_amd64.whl
whispercpp --help

The executable and DLLs are installed under site-packages/whispercpp_runtime/bin. Pip adds the whispercpp command to the environment's Scripts directory. Python applications can locate the executable with whispercpp_runtime.binary_path().

Standalone ZIP

Download whispercpp-v1.8.5-windows-x64.zip from the release page and extract it. Keep whisper-cli.exe and all included DLLs together. Python is not required.

Release assets include SHA256SUMS.txt for checksum verification.

Usage

Download a compatible GGML model using the upstream model instructions, then run:

whispercpp -m "ggml-small.bin" -f "audio.wav" -l en -fa

For the standalone ZIP, replace whispercpp with .\whisper-cli.exe. Use --help to list available options.

Building from source

The current package, 1.8.5.1, uses unmodified upstream commit f24588a272ae8e23280d9c220536437164e6ed28, which reports whisper.cpp version 1.8.5.

Build with MSVC v143, CMake, Ninja, and the Vulkan SDK. Open an x64 Native Tools command prompt and use a short build path:

git clone https://github.com/ggml-org/whisper.cpp.git source
git -C source checkout --detach f24588a272ae8e23280d9c220536437164e6ed28
cmake -S source -B build -G Ninja -DCMAKE_BUILD_TYPE=Release -DCMAKE_C_COMPILER=cl -DCMAKE_CXX_COMPILER=cl -DBUILD_SHARED_LIBS=ON -DGGML_VULKAN=ON -DGGML_OPENMP=ON -DGGML_NATIVE=ON -DWHISPER_BUILD_TESTS=OFF
cmake --build build --target whisper-cli -j 4

Native CPU tuning targets the build machine. Check the resulting instruction-set requirements before distributing the build.

To package a wheel, place the executable and DLLs in src/whispercpp_runtime/bin, along with the upstream license as LICENSE-whisper.cpp.txt and a BUILD.json describing the source revision and build settings. Then run:

python -m pip wheel . --no-deps -w dist

License and provenance

This is a community distribution of whisper.cpp. Runtime packages include the upstream license and build metadata. Preserve those files when redistributing the runtime.