hipRAFT documentation#
2025-10-30
3 min read time
hipRAFT is a library of fundamental algorithms and primitives for machine-learning and data-mining workloads that can be run on AMD GPUs. It is part of the AMD ROCm Data Science toolkit (ROCm-DS), an open-source software collection for high-performance data science applications. Forked from the NVIDIA® RAPIDS® RAFT project, hipRAFT brings the same rich functionality to the HIP/ROCm stack while preserving the directory structure, file naming, and API naming, to minimize porting friction for developers using both projects. It offers both a modern C++ interface for systems developers and fully featured Python bindings for rapid prototyping and data-science workflows. For more information, see What is hipRAFT?
Key highlights in hipRAFT v0.1.0 include:
Full integration with the ROCm-DS ecosystem – Serves as the computational foundation for hipGRAPH and hipVS, providing shared GPU-accelerated primitives across all components.
Expanded C++ and Python support – Offers a modern C++ interface for system-level integration and Python bindings for rapid prototyping and data-science workflows.
Rich module coverage – Delivers accelerated functionality across core domains, including:
Linear Algebra – GPU-optimized BLAS routines, solvers, and matrix factorizations.
Matrix Utilities – Advanced arithmetic, manipulation, and reduction operations.
Sparse Operations – Efficient computation and solvers for large-scale sparse data.
Label Processing – High-performance label extraction and remapping utilities for ML and graph algorithms.
Optimization Solvers – Support for linear assignment (Hungarian algorithm) and minimum spanning tree (MST) solvers.
Random Number Generation – GPU-based random utilities for ML training and simulation.
Performance and resource management improvements – Enhanced GPU memory handling, logging, and profiling utilities for efficient large-scale computation.
The hipRAFT code is open and hosted at ROCm-DS/hipRaft.
To contribute to the documentation refer to Contributing to ROCm-DS.
You can find licensing information on the Licensing page.