Install from Source

Install from Source#

This page gives instructions on how to build and install the TVM package from source.

Step 1. Install Dependencies#

Apache TVM requires the following dependencies:

  • CMake (>= 3.24.0)

  • LLVM (recommended >= 15)

  • Git

  • A recent C++ compiler supporting C++ 20, at the minimum

    • GCC 10

    • Clang 10

    • Apple Clang 14

    • Visual Studio 2022

    Optional dependencies that use newer C++20 standard library facilities, such as std::format, may require a newer standard library (for example GCC 13 or newer on Linux).

  • Python (>= 3.10)

  • (Optional) Conda (Strongly Recommended)

System Dependencies (Non-Conda)#

If you are not using Conda, TVM requires several system libraries. On Ubuntu/Debian systems, install them with:

sudo apt update
sudo apt install zlib1g-dev libxml2-dev

For other operating systems, please refer to your package manager documentation.

The easiest way to manage dependency is via conda, which maintains a set of toolchains including LLVM across platforms. To create the environment of those build dependencies, one may simply use:

# make sure to start with a fresh environment
conda env remove -n tvm-build-venv
# create the conda environment with build dependency
conda create -n tvm-build-venv -c conda-forge \
    "llvmdev>=15" \
    "cmake>=3.26" \
    git \
    python=3.11
# enter the build environment
conda activate tvm-build-venv

Note

For Frontend Contributors (TFLite): If you plan to run or contribute to the frontend tests (e.g., test_frontend_tflite.py), you must install tensorflow==2.19.0.

pip install tensorflow==2.19.0

Python 3.11 is supported.

Step 2. Get Source from GitHub#

You can also choose to clone the source repo from GitHub.

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

TVM pins the standalone Apache TIRx compiler and core runtime in 3rdparty/tvm-tirx. Recursive submodule initialization also checks out its FFI dependency. After updating TVM, run git submodule update --init --recursive to restore the pinned versions.

Note

It’s important to use the --recursive flag when cloning the TVM repository, which will automatically clone the submodules. If you forget to use this flag, you can manually clone the submodules by running git submodule update --init --recursive in the root directory of the TVM repository.

Step 3. Configure and Build#

Configure the build from the repository root with explicit CMake options. Enable GPU SDKs and vendor libraries as needed:

cd tvm
cmake -S . -B build \
    -DCMAKE_BUILD_TYPE=RelWithDebInfo \
    -DTVM_TIRX_USE_LLVM="llvm-config --ignore-libllvm --link-static" \
    -DTVM_HIDE_PRIVATE_SYMBOLS=ON \
    -DTVM_TIRX_USE_CUDA=OFF \
    -DTVM_TIRX_USE_METAL=OFF \
    -DTVM_TIRX_USE_VULKAN=OFF \
    -DTVM_TIRX_USE_OPENCL=OFF \
    -DTVM_USE_CUBLAS=OFF \
    -DTVM_USE_CUDNN=OFF \
    -DTVM_USE_CUTLASS=OFF

CMake stores these values in build/CMakeCache.txt. To change an option, rerun the configure command with its new -D value. Parent TVM features use TVM_USE_*; core features retain TVM_TIRX_USE_* and FFI options retain TVM_FFI_*. CMake also supports local CMakeUserPresets.json presets.

Note

TVM_HIDE_PRIVATE_SYMBOLS is a configuration option that enables the -fvisibility=hidden flag. This flag helps prevent potential symbol conflicts between TVM and PyTorch. These conflicts arise due to the frameworks shipping LLVMs of different versions.

CMAKE_BUILD_TYPE controls default compilation flag:

  • Debug sets -O0 -g

  • RelWithDebInfo sets -O2 -g -DNDEBUG (recommended)

  • Release sets -O3 -DNDEBUG

After configuration succeeds, build with:

cmake --build build --parallel $(nproc)

Note

nproc may not be available on all systems, please replace it with the number of cores on your system

A successful build produces the compiler and runtime libraries under build/lib/.

Apache TVM relies on the tvm-ffi package to support its python bindings. Therefore, after we finish the build, we need to install the tvm-ffi package.

pip install ./3rdparty/tvm-tirx/3rdparty/tvm-ffi

Leaving the build environment tvm-build-venv, there are two ways to install the successful build into your environment:

  • Install via environment variable

export TVM_HOME=/path-to-tvm
pip install --target=$TVM_HOME/python $TVM_HOME/3rdparty/tvm-tirx/3rdparty/tvm-ffi
export PYTHONPATH=$TVM_HOME/3rdparty/tvm-tirx/python:$TVM_HOME/python:$PYTHONPATH
export TVM_TIRX_LIBRARY_PATH=$TVM_HOME/build/lib

  • Install via pip local project

conda activate your-own-env
conda install python # make sure python is installed
pip install "scikit-build-core>=0.11" "setuptools-scm>=8"
pip install /path-to-tvm/3rdparty/tvm-tirx
pip install --no-build-isolation /path-to-tvm

Step 4. Validate Installation#

Using a compiler infrastructure with multiple language bindings could be error-prone. Therefore, it is highly recommended to validate Apache TVM installation before use.

Step 1. Locate TVM Python package. The following command can help confirm that TVM is properly installed as a python package and provide the location of the TVM python package:

>>> python -c "import tvm; print(tvm.__file__)"
/some-path/lib/python3.11/site-packages/tvm/__init__.py

Step 2. Confirm which TVM library is used. When maintaining multiple build or installation of TVM, it becomes important to double check if the python package is using the proper libtvm with the following command:

>>> python -c "import tvm; print(tvm.base._LOADED_LIBS['tvm_runtime'])"
<CDLL '/some-path/lib/python3.11/site-packages/tvm/libtvm_runtime.dylib', handle 95ada510 at 0x1030e4e50>

Step 3. Inspect the build configuration. For a local source build, inspect the CMake cache to check which options were configured:

cmake -LA -N build

The output includes parent TVM_USE_* options and the core’s TVM_TIRX_USE_* options. Keep build/CMakeCache.txt when reporting a configuration problem.

Step 4. Check device detection. Sometimes it could be helpful to understand if TVM could detect your device at all with the following commands:

>>> python -c "import tvm; print(tvm.metal().exist)"
True # or False
>>> python -c "import tvm; print(tvm.cuda().exist)"
False # or True
>>> python -c "import tvm; print(tvm.vulkan().exist)"
False # or True

Please note that the commands above verify the presence of an actual device on the local machine for the TVM runtime (not the compiler) to execute properly. However, TVM compiler can perform compilation tasks without requiring a physical device. As long as the necessary toolchain, such as NVCC, is available, TVM supports cross-compilation even in the absence of an actual device.

Step 5. Extra Python Dependencies#

Building from source does not ensure the installation of all necessary Python dependencies.

Python Dependencies#

The following commands can be used to install the extra Python dependencies:

  • Necessary dependencies:

pip3 install numpy cython

  • If you want to use RPC Tracker

pip3 install tornado

  • If you want to use auto-tuning module

pip3 install tornado psutil 'xgboost>=1.1.0' cloudpickle

Windows-Specific Build Notes#

If you’re building TVM on Windows, note these platform-specific considerations:

Path Conventions#

  • Use forward slashes (/) in Python/CMake paths, not Windows backslashes

  • Example: -DCMAKE_TOOLCHAIN_FILE=C:/toolchains/windows.cmake.

Advanced Build Configuration#

Ccache#

On supported platforms, the Ccache compiler wrapper may be helpful for reducing TVM’s build time, especially when building with cutlass. There are several ways to enable CCache in TVM builds:

  • Configure with -DTVM_USE_CCACHE=AUTO (the default). CCache will be used if it is found.

  • Ccache’s Masquerade mode. This is typically enabled during the Ccache installation process. To have TVM use Ccache in masquerade, simply specify the appropriate C/C++ compiler paths when configuring TVM’s build system. For example: cmake -DCMAKE_CXX_COMPILER=/usr/lib/ccache/c++ ....

  • Ccache as CMake’s C++ compiler prefix. When configuring TVM’s build system, set the CMake variable CMAKE_CXX_COMPILER_LAUNCHER to an appropriate value. E.g. cmake -DCMAKE_CXX_COMPILER_LAUNCHER=ccache ....

Building on Windows#

TVM support build via MSVC using CMake. You will need to obtain a visual studio compiler. The minimum required VS version is Visual Studio Enterprise 2019 (NOTE: we test against GitHub Actions’ windows-latest Runner, so see that page for full details. We recommend following Step 1. Install Dependencies to obtain necessary dependencies and get an activated tvm-build environment. Then you can run the following command to build

mkdir build
cd build
cmake ..
cd ..

The above command generates the solution file under the build directory. You can then run the following command to build

cmake --build build --config Release -- /m

CUDA Configuration#

For CUDA support on Windows:

set CUDA_PATH=C:\Program Files\NVIDIA GPU Computing Toolkit\CUDA\v11.8
set PATH=%CUDA_PATH%\bin;%PATH%
cmake .. -DTVM_TIRX_USE_CUDA=ON

CMake & Compiler Setup#

  • Specify generator: cmake -G "Visual Studio 16 2019" -A x64 ..

  • Ensure Python is in PATH or specify: -DPython_EXECUTABLE=C:\Python311\python.exe

Building ROCm support#

Currently, ROCm is supported only on linux, so all the instructions are written with linux in mind.

  • Configure with -DTVM_TIRX_USE_ROCM=ON or set it to the ROCm SDK path.

  • You need to first install HIP runtime from ROCm. Make sure the installation system has ROCm installed in it.

  • Install LLVM (>= 15), and LLD, make sure ld.lld is available via command line.

Enable C++ Tests#

We use Google Test to drive the C++ tests in TVM. The easiest way to install GTest is from source.

git clone https://github.com/google/googletest
cd googletest
mkdir build
cd build
cmake -DBUILD_SHARED_LIBS=ON ..
make
sudo make install

After installing GTest, the C++ tests can be built and started with ./tests/scripts/task_cpp_unittest.sh, or built via CMake with -DTVM_USE_GTEST=ON and then running ./build/cpptest.

Contents