tenferro-rs: A New Tensor Stack for Rust Scientific Computing Launched

Introduction to tenferro-rs

tenferro-rs is a new Rust-native dense tensor stack designed specifically for scientific computing. It incorporates features such as linear algebra, PyTorch-style eager autodiff, JAX-style traced transforms, and NumPy-style einsum support, as well as extensible operation crates for various computational needs.

Transition from Julia to Rust

Historically, most tensor-network code, including that used by the tensor4all team, was developed in Julia. However, as projects grew, challenges such as type instability and lengthy compile times became prominent, prompting a shift towards Rust for its performance benefits.

The Need for a Unified Tensor Library

As the development of the tensor stack in Rust began, it became clear that existing libraries were fragmented. While Rust has several libraries for specific tasks—like ndarray for arrays and Burn for deep learning—there was a lack of a comprehensive tensor layer capable of integrating features like autodiff and einsum, essential for scientific applications. tenferro-rs aims to fill this gap.

Integration with the Rust Ecosystem

The creation of tenferro-rs builds on various existing Rust libraries such as faer for linear algebra and CubeCL for GPU kernels. Rather than starting from scratch, the development team focused on enhancing and integrating these existing resources to create a functional tensor stack, which includes column-major storage, dynamic shapes, and both CPU and CUDA compatibility.

Implications for Scientific Computing

The launch of tenferro-rs represents a significant development in the Rust programming ecosystem, particularly for scientific computing. By providing a well-rounded tensor stack, it supports more efficient computations and code organization for large-scale projects, which may lead to broader adoption of Rust in the scientific community.