rstensor

rstensor is an open-source library for tensor computations written in Rust. It targets scientific computing and machine learning workloads, offering a safe, high-performance core for multi-dimensional array operations. The project emphasizes ergonomic APIs, interoperability with other ecosystems, and a pluggable backend architecture.

The library provides an N-dimensional tensor type with support for dynamic and static shapes, broadcasting, indexing, slicing, and reductions. It includes an automatic differentiation engine suitable for prototyping neural networks, as well as a growing set of linear algebra primitives such as matmul, norms, and trace. Element-wise operations are defined with Rust's type system to minimize runtime overhead.

rstensor uses a modular backend design. The CPU backend emphasizes multi-threaded execution through Rayon and memory-safe parallelism. GPU acceleration is possible via pluggable backends, including Vulkan- or CUDA-based compute paths, enabling offloading of kernels when available. The library also supports memory layout options and strides to optimize performance for various workloads.

Interoperability with other languages is supported through bindings and data interchange formats. It offers reading and writing of NumPy-compatible formats (.npy and .npz) and provides lightweight bindings to Python via PyO3 for researchers who want to use rstensor alongside Python tools.

rstensor is designed with safety and portability in mind, leveraging Rust's ownership model to prevent common memory errors. The codebase emphasizes modularity, minimal dependencies, and comprehensive documentation to facilitate contributions and experimentation.

The project is hosted on a public repository with an open-source license, active issue trackers, and community-maintained examples and tutorials. It aims to complement existing Rust and data-science ecosystems by providing a robust core for tensor computations.