Installation guide | Avelang Documentation

Prerequisites

LLVM, MLIR, and Clang 22. For AMDGPU target we recommend using the ROCm 7.2.x fork of LLVM.
CMake 3.25 or newer
Ninja
Python 3.10 or newer
CUDA for NVIDIA targets, or ROCm for AMDGPU targets

Building and installing Ave

We recommend building Ave inside a uv Python virtual environment, and using Clang for faster compilation. Ave supports both CUDA and ROCm backends through the AVE_LANG_BACKEND CMake option. Valid values are cuda, rocm, or cuda;rocm. For example, to enable the CUDA backend, run the following commands from the source directory (replace /path/to/llvm with your LLVM install prefix, such as /usr/local in our CUDA environment):

uv venv .venv
source .venv/bin/activate
CMAKE_ARGS='-DAVE_LANG_BACKEND=cuda -DWITH_PYTHON=ON \
  -DCMAKE_C_COMPILER=/path/to/llvm22/bin/clang \
  -DCMAKE_CXX_COMPILER=/path/to/llvm22/bin/clang++ \
  -DCMAKE_PREFIX_PATH=/path/to/llvm22' uv pip install -e .

For ROCm:

uv venv .venv
source .venv/bin/activate
CMAKE_ARGS="-DAVE_LANG_BACKEND=rocm -DWITH_PYTHON=ON \
  -DCMAKE_C_COMPILER=/path/to/llvm22/bin/clang \
  -DCMAKE_CXX_COMPILER=/path/to/llvm22/bin/clang++ \
  -DCMAKE_PREFIX_PATH='/path/to/llvm22;/opt/rocm'" uv pip install -e .

You should be able to run the python test via python -m pytest test once the build and installation are completed.

Development build

To avoid rebuilding the full package on every pip install during development, configure a separate build directory and write the Python extension directly into python/. For a CUDA development build:

uv venv .venv
source .venv/bin/activate
uv pip install -e ".[dev]"
cmake -S . -B build -GNinja -DCMAKE_BUILD_TYPE=Release \
  -DAVE_LANG_BACKEND=cuda \
  -DWITH_PYTHON=ON \
  -DCMAKE_C_COMPILER=/path/to/llvm/bin/clang \
  -DCMAKE_CXX_COMPILER=/path/to/llvm/bin/clang++ \
  -DCMAKE_PREFIX_PATH=/path/to/llvm \
  -DAVE_LANG_PYTHON_LIBRARY_OUTPUT_DIRECTORY="$(pwd)/python"
ninja -C build
ctest --test-dir build --output-on-failure
PYTHONPATH=python python -m pytest test

For ROCm, switch AVE_LANG_BACKEND to rocm, set CMAKE_PREFIX_PATH to '/path/to/llvm;/opt/rocm', and run PYTHONPATH=python python -m pytest test.

Running ninja -C build enables incremental builds of Ave to speed up development.

After installation, continue with the tutorials to write and launch Ave kernels.