Build from Source

Build from Source#

This guide provides instructions on how to build the TIRx package from source. uv sync creates the project .venv and installs the package in editable mode.

Prerequisite

  • Python: 3.10 or newer.

  • Compiler: a C++20-capable GCC, Clang, or MSVC toolchain.

  • Build tools: Git and uv.

Build the Python Package#

Download the source via:

git clone --recursive https://github.com/apache/tvm-tirx
cd tvm-tirx

Note

Always clone with --recursive to pull submodules. If you already cloned without it, run:

git submodule update --init --recursive

Build the Python package with scikit-build-core, which drives CMake to compile the C++ libraries and supplies CMake/Ninja when needed in the isolated build:

uv sync --no-default-groups

The editable install reflects Python-only changes immediately without rebuilding.

CMake flags can be exported through CMAKE_ARGS before running the install command. For example, to switch off the CUDA runtime (which is automatically switched on when available):

export CMAKE_ARGS="-DTVM_TIRX_USE_CUDA=OFF"

Available CMake options (see CMakeLists.txt) include:

  • TVM_TIRX_USE_LLVM – Prefer the build Python’s static LLVM SDK, or select an explicit LLVM_DIR / llvm-config command (default: ON). OFF disables LLVM.

  • TVM_TIRX_USE_CUDA – Enable the CUDA runtime (default: AUTO).

  • TVM_TIRX_USE_METAL – Enable the Metal runtime on macOS (default: OFF).

  • TVM_TIRX_USE_Z3 – Enable the Z3 solver (default: AUTO).

  • TVM_TIRX_USE_RUST – Build the operators written in Rust (rust/) with cargo (default: AUTO, which needs a cargo as new as the crate’s rust-version and the crate’s initialized TVM FFI submodule, with other dependencies from crates.io or cargo’s cache).

  • TVM_TIRX_BUILD_TESTS – Build C++ tests (default: OFF).

Verify the installation:

uv run --no-sync python -c "import tvm_tirx; print(tvm_tirx.__version__)"
uv run --no-sync tvm-tirx-config -h

Tip

Use tvm-tirx-config to query include and link flags for external C++ projects:

uv run --no-sync tvm-tirx-config --includedir
uv run --no-sync tvm-tirx-config --libfiles

On Linux/macOS, query linker flags for the compiler or runtime libraries:

uv run --no-sync tvm-tirx-config --libs compiler      # Compiler and its dependencies
uv run --no-sync tvm-tirx-config --libs runtime       # Core runtime
uv run --no-sync tvm-tirx-config --libs runtime cuda  # Core and CUDA runtimes (CUDA build)
uv run --no-sync tvm-tirx-config --ldflags            # Library search paths

Build with CMake#

Install only the build tools and SDKs, without installing the Python package, then configure and build:

uv sync --only-group build
uv run --no-sync cmake -S . -B build_cpp -G Ninja -DCMAKE_BUILD_TYPE=Release
uv run --no-sync cmake --build build_cpp --parallel

Build Wheels with Customized LLVM#

By default, builds use the pinned prebuilt LLVM/Z3 static SDK packages from PyPI specified in pyproject.toml; no manual SDK setup is needed.

An explicit LLVM_DIR or TVM_TIRX_USE_LLVM=/path/to/llvm-config selects an external LLVM installation. Supply its development libraries and transitive dependencies before building.

Note

When building wheels for release, we recommend linking LLVM statically into the wheel to avoid conflicts with LLVM libraries used by other packages. Pass --link-static to llvm-config to enable this.

CMAKE_ARGS="-DTVM_TIRX_USE_LLVM='/path/to/llvm-config --link-static'" uv build --wheel

For macOS wheels using the default SDK, set MACOSX_DEPLOYMENT_TARGET=14.0. See Reproduce CI/CD for the publication toolchain configuration.

Troubleshooting#

  • Rebuilding after C++ changes: re-run uv sync --no-default-groups --reinstall-package apache-tvm-tirx. Editable installs automatically reflect only Python changes.

  • Submodules missing: run git submodule update --init --recursive from the repository root.

See also

  • Reproduce CI/CD: Reproduce linters, unit tests, and wheel builds locally.