RAFT

RAFT is an open-source suite of GPU-accelerated libraries providing fundamental algorithms and primitives for machine learning and information retrieval. It is designed as a C++ header-only template library (with optional shared library), and also offers lightweight Python wrappers (pylibraft, raft-dask).

Features

• CUDA-accelerated primitives for machine learning and data mining
• Data Formats: Support for both sparse and dense data, conversions, and data generation
• Dense Operations: Linear algebra, matrix and vector operations, reductions, slicing, norms, factorization, least squares, SVD, and eigenvalue problems
• Sparse Operations: Linear algebra, eigenvalue problems, slicing, norms, reductions, factorization, symmetrization, components, and labeling
• Solvers: Combinatorial optimization and iterative solvers
• Statistics: Sampling, moments and summary statistics, metrics, and model evaluation
• Tools & Utilities: Common tools for CUDA application development, multi-node/multi-GPU infrastructure
• Memory Management: Integrates with RAPIDS Memory Manager (RMM) for flexible memory allocation strategies
• Multi-dimensional Arrays: APIs use mdspan for multi-dimensional array views, similar to NumPy's ndarray, and provide mdarray for memory management on host and device (GPU)
• Python Bindings: pylibraft and raft-dask for Python integration

Notes

• RAFT is intended for use by developers building high-performance, GPU-accelerated applications, not directly for exploratory data science.
• Some vector search and clustering algorithms have migrated to cuVS, and their RAFT implementations will be deprecated in future releases.

Pricing

RAFT is open-source software and is available for free under its respective open-source license.