Tensor Computing Interface: An Application-Oriented, Lightweight Interface for Portable High-Performance Tensor Network Applications
Tensor networks (TNs) are a central computational tool in quantum science and artificial intelligence. However, the lack of unified software interface across tensor-computing frameworks severely limits the portability of TN applications, coupling algorithmic development to specific hardware and software back ends. To address this challenge, we introduce the Tensor Computing Interface (TCI) -- an application-oriented, lightweight application programming interface designed to enable framework-independent, high-performance TN applications. TCI provides a well-defined type system that abstracts tensor objects together with a minimal yet expressive set of core functions covering essential tensor manipulations and tensor linear-algebra operations. Through numerical demonstrations on representative tensor-network applications, we show that codes written against TCI can be migrated seamlessly across heterogeneous hardware and software platforms while achieving performance comparable to native framework implementations. We further release an open-source implementation of TCI based on \textit{Cytnx}, demonstrating its practicality and ease of integration with existing tensor-computing frameworks.