Heat is a distributed tensor framework for high performance data analytics.

Project Status

Table of Contents

• What is Heat for?
• Features
• Getting Started
• Installation
  • Requirements
  • pip
  • conda
• Support Channels
• Contribution guidelines
  • Resources
• License
• Citing Heat
• FAQ
• Acknowledgements

What is Heat for?

Heat builds on PyTorch and mpi4py to provide high-performance computing infrastructure for memory-intensive applications within the NumPy/SciPy ecosystem.

With Heat you can: - port existing NumPy/SciPy code from single-CPU to multi-node clusters with minimal coding effort; - exploit the entire, cumulative RAM of your many nodes for memory-intensive operations and algorithms; - run your NumPy/SciPy code on GPUs (CUDA, ROCm, coming up: Apple MPS).

For a example that highlights the benefits of multi-node parallelism, hardware acceleration, and how easy this can be done with the help of Heat, see, e.g., our blog post on trucated SVD of a 200GB data set.

Check out our coverage tables to see which NumPy, SciPy, scikit-learn functions are already supported.

If you need a functionality that is not yet supported: - search existing issues and make sure to leave a comment if someone else already requested it; - open a new issue.

Check out our features and the Heat API Reference for a complete list of functionalities.

Features

• High-performance n-dimensional arrays
• CPU, GPU, and distributed computation using MPI
• Powerful data analytics and machine learning methods
• Seamless integration with the NumPy/SciPy ecosystem
• Python array API (work in progress)

Getting Started

Go to Quick Start for a quick overview. For more details, see Installation.

You can test your setup by running the heat_test.py script:

shell mpirun -n 2 python heat_test.py

It should print something like this:

shell x is distributed: True Global DNDarray x: DNDarray([0, 1, 2, 3, 4, 5, 6, 7, 8, 9], dtype=ht.int32, device=cpu:0, split=0) Global DNDarray x: Local torch tensor on rank 0 : tensor([0, 1, 2, 3, 4], dtype=torch.int32) Local torch tensor on rank 1 : tensor([5, 6, 7, 8, 9], dtype=torch.int32)

Check out our Jupyter Notebook Tutorials, choose local to try things out on your machine, or hpc if you have access to an HPC system.

The complete documentation of the latest version is always deployed on Read the Docs.

Installation

Requirements

Basics

• python >= 3.10
• MPI (OpenMPI, MPICH, Intel MPI, etc.)
• mpi4py >= 3.0.0
• pytorch >= 2.0.0

Parallel I/O

• h5py
• netCDF4

GPU support

In order to do computations on your GPU(s): - your CUDA or ROCm installation must match your hardware and its drivers; - your PyTorch installation must be compiled with CUDA/ROCm support.

HPC systems

On most HPC-systems you will not be able to install/compile MPI or CUDA/ROCm yourself. Instead, you will most likely need to load a pre-installed MPI and/or CUDA/ROCm module from the module system. Maybe, you will even find PyTorch, h5py, or mpi4py as (part of) such a module. Note that for optimal performance on GPU, you need to usa an MPI library that has been compiled with CUDA/ROCm support (e.g., so-called "CUDA-aware MPI").

pip

Install the latest version with

bash pip install heat[hdf5,netcdf] where the part in brackets is a list of optional dependencies. You can omit it, if you do not need HDF5 or NetCDF support.

conda

The conda build includes all dependencies including OpenMPI. bash conda install -c conda-forge heat

Support Channels

Go ahead and ask questions on GitHub Discussions. If you found a bug or are missing a feature, then please file a new issue. You can also get in touch with us on Mattermost (sign up with your GitHub credentials). Once you log in, you can introduce yourself on the Town Square channel.

Contribution guidelines

We welcome contributions from the community, if you want to contribute to Heat, be sure to review the Contribution Guidelines and Resources before getting started!

We use GitHub issues for tracking requests and bugs, please see Discussions for general questions and discussion. You can also get in touch with us on Mattermost (sign up with your GitHub credentials). Once you log in, you can introduce yourself on the Town Square channel.

If you’re unsure where to start or how your skills fit in, reach out! You can ask us here on GitHub, by leaving a comment on a relevant issue that is already open.

If you are new to contributing to open source, this guide helps explain why, what, and how to get involved.

Resources

• Heat Tutorials
• Heat API Reference

Parallel Computing and MPI:

• David Henty's course
• Wes Kendall's Tutorials
• Rolf Rabenseifner's MPI course material (including C, Fortran and Python via mpi4py)

mpi4py

• mpi4py docs
• Tutorial # License

Heat is distributed under the MIT license, see our LICENSE file.

Citing Heat

If Heat contributed to a publication, please cite our main paper.

Preferred Citation:

Götz, M., Debus, C., Coquelin, D., et al. (2020). HeAT - a Distributed and GPU-accelerated Tensor Framework for Data Analytics. In 2020 IEEE International Conference on Big Data (Big Data) (pp. 276-287). IEEE. DOI: 10.1109/BigData50022.2020.9378050.

bibtex @inproceedings{heat2020, title={{HeAT -- a Distributed and GPU-accelerated Tensor Framework for Data Analytics}}, author={Markus Götz and Charlotte Debus and Daniel Coquelin and Kai Krajsek and Claudia Comito and Philipp Knechtges and Björn Hagemeier and Michael Tarnawa and Simon Hanselmann and Martin Siggel and Achim Basermann and Achim Streit}, booktitle={2020 IEEE International Conference on Big Data (Big Data)}, year={2020}, pages={276-287}, publisher={IEEE}, doi={10.1109/BigData50022.2020.9378050} }

Other Relevant Publications

For the RSE perspective and latest benchmarks:

Hoppe, F., et al. (2025). Engineering a large-scale data analytics and array computing library for research: Heat. Electronic Communications of the EASST, 83.

bibtex @article{heat2025rse, title={Engineering a large-scale data analytics and array computing library for research: Heat}, volume={83}, url={[https://eceasst.org/index.php/eceasst/article/view/2626](https://eceasst.org/index.php/eceasst/article/view/2626)}, DOI={10.14279/eceasst.v83.2626}, journal={Electronic Communications of the EASST}, author={Hoppe, Fabian and Gutiérrez Hermosillo Muriedas, Juan Pedro and Tarnawa, Michael and Knechtges, Philipp and Hagemeier, Björn and Krajsek, Kai and Rüttgers, Alexander and Götz, Markus and Comito, Claudia}, year={2025} }

For the neural networks module (DASO):

Coquelin, D., et al. (2022). Accelerating neural network training with distributed asynchronous and selective optimization (DASO). J Big Data 9, 14.

bibtex @Article{DASO2022, author={Coquelin, Daniel and Debus, Charlotte and G{\"o}tz, Markus and von der Lehr, Fabrice and Kahn, James and Siggel, Martin and Streit, Achim}, title={Accelerating neural network training with distributed asynchronous and selective optimization (DASO)}, journal={Journal of Big Data}, year={2022}, volume={9}, number={1}, pages={14}, doi={10.1186/s40537-021-00556-1} }

For specific software versions: Please use the Zenodo DOI provided with each release.

FAQ

Work in progress...

<!-- - Users - Developers - Students - system administrators -->

Acknowledgements

This work is supported by the Helmholtz Association Initiative and Networking Fund under project number ZT-I-0003 and the Helmholtz AI platform grant.

This project has received funding from Google Summer of Code (GSoC) in 2022.

This work is partially carried out under a programme of, and funded by, the European Space Agency. Any view expressed in this repository or related publications can in no way be taken to reflect the official opinion of the European Space Agency.