Building and installing hipMM from source

Building and installing hipMM from source#

2025-11-14

4 min read time

Applies to Linux

For developers, the following topic provides details on building the C++ components, running tests and benchmarks, and for building the full hipMM installation including the Python package. End users should see the Installation instructions.

Build procedure for C++ components#

Build the hipMM C++/HIP components using the following command:

./build.sh lib tests benchmarks

Here, tests and benchmarks are optional flags that enable the respective functionality.

Note

To fetch the dependencies, git needs to be installed on your system.

Running tests and benchmarks#

To run the tests use:

ctest --test-dir cpp/build/

To run the benchmarks, use:

find cpp/build/benchmarks/ -type f -executable -exec {} \;

Building and installing hipMM including the Python layer#

You will perform the following steps:

  1. Install Conda

  2. Clone the hipMM repository

  3. Create and activate hipMM Conda environment

  4. Build and install hipMM

  5. Verify correctness of installation

Step 1: Install conda#

hipMM must be built inside of a predefined Conda environment to ensure that it is working properly. While the hipMM build process fetches C++ dependencies itself, it has Cython and Python dependencies (CuPy, Numba HIP, hipMM, HIP Python, ROCm LLVM Python) that need to be installed into the hipMM Conda environment before you can build and use the package.

On an x86 Linux machine, you can download and install Miniforge via:

wget https://github.com/conda-forge/miniforge/releases/latest/download/Miniforge3-Linux-x86_64.sh
sh Miniforge3-Linux-x86_64.sh

For other architectures and operating systems take a look at the webpage of Miniforge.

Step 2: Clone the hipMM repository#

Create a work directory /tmp/hipmm and clone hipMM into this repository:

mkdir -p /tmp/hipmm # NOTE: feel free to adapt

cd /tmp/hipmm
git clone https://github.com/ROCm-DS/hipMM hipmm -b release/rocmds-25.10

Step 3: Create and activate the hipMM conda environment#

Given the following conda environment (conda/environments/build-release_rocm-latest_arch-x86_64.yaml):

channels:
- rapidsai
- rapidsai-nightly
- conda-forge
dependencies:
- c-compiler
- clang-tools==16.0.6
- clang==16.0.6
- cmake>=3.26.4,<4,!=3.30.0
- cxx-compiler
- cython>=3.0.0
- fmt>=11.0.2,<12
- gcc_linux-64=11.*
- gcovr>=5.0
- identify>=2.5.20
- ipython
- make
- ninja
- numpy>=1.23,<3.0a0
- pre-commit
- pytest
- pytest-cov
- python>=3.10,<3.13
- rapids-build-backend>=0.3.0,<0.4.0.dev0
- scikit-build-core >=0.10.0
- spdlog>=1.14.1,<1.15
- pip
- pip:
- --pre
- --extra-index-url=https://pypi.amd.com/rocm-7.0.2/simple
- rocm-llvm-python~=7.0.2.0
- hip-python~=7.0.2.0
- hip-python-as-cuda~=7.0.2.0
- numba-hip~=0.1.3
- amd-cupy~=13.5.1
name: hipmm_dev

Create and activate the hipmm_dev Conda environment via:

cd /tmp/hipmm/hipmm

conda env create --name hipmm_dev --file conda/environments/build-release_rocm-latest_arch-x86_64.yaml
conda activate hipmm_dev

Step 4: Build and install hipMM#

The build.sh command creates the build directory in the cpp subfolder of the cloned hipMM git repository. You can build and install librmm and rmm separately, and you can also build without installing using the -n option. Use build.sh -h to display the help text for the script.

(hipmm_dev) $ export CXX="hipcc"    # Cython CXX compiler, adjust according to your setup.
(hipmm_dev) $ export CMAKE_PREFIX_PATH="${CMAKE_PREFIX_PATH}:/opt/rocm/lib/cmake"     # Locate ROCm CMake packages
(hipmm_dev) $ ./build.sh librmm rmm     # Build and install librmm and rmm (can also use the default ./build.sh)

Note

When rebuilding it is recommended to remove previous build files. When you are using the ./build.sh script, this can be accomplished by additionally specifying clean. For example: ./build.sh clean rmm.

Step 5: Verify correct installation#

After building and installing hipMM for use in the Conda environment hipmm_dev, you can verify the correctness of the installation as follows:

conda activate hipmm_dev

python3

Then enter the following Python code commands:

import hipmm

print(hipmm.__version__)

You should see output similar to:

Python 3.12.11 | packaged by conda-forge | (main, Jun  4 2025, 14:45:31) [GCC 13.3.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import hipmm
>>> print(hipmm.__version__)
'3.0.00'