Install SparseTIR

Currently we only support build SparseTIR from source code, you’ll need to build the shared library (which is written in C++) first and then install Python bindings. It’s worth noting that SparseTIR is a fork of Apache TVM project, and you don’t need to install Apache TVM to use Sparse TIR.


We recommend user to install following packages before compiling SparseTIR shared library:

  • A recent C++ compiler supporting C++ 14.

  • CMake 3.18 or higher

  • LLVM 10 or higher

  • CUDA Toolkit 11 or higher

  • Python 3.9 or higher

Build the Shared Library

The first step is to compile source code written in C++.

git clone --recursive
echo set\(USE_LLVM \"llvm-config --ignore-libllvm --link-static\"\) >> config.cmake
echo set\(HIDE_PRIVATE_SYMBOLS ON\) >> config.cmake
echo set\(USE_CUDA ON\) >> config.cmake
echo set\(USE_CUBLAS ON\) >> config.cmake
echo set\(USE_CUDNN ON\) >> config.cmake
mkdir -p build
cd build
cmake ..
make -j$(nproc)

Install Python Binding

If compilation is successful, the next step is to install SparseTIR binding for Python, you can either install Python package via:

cd python
python3 install

or set environment variable ${PYTHONPATH}:

export SPARSETIR_PATH=$(pwd)