SHI Collaboration Profiles

Profile pages for Sustainable Horizons Institute SRP 2025-2026 Project Leaders


T

Tianyi Shi

Lawrence Berkeley National Laboratory

Biography

Tianyi Shi is a postdoctoral scholar at Lawrence Berkeley National Laboratory in the Scalable Solvers Group within the Computing Sciences Area. He received his Ph.D. in Applied Mathematics from Cornell University. Tianyi’s research focuses on developing efficient and scalable algorithms for sparse and data-sparse matrices and tensors, with applications in computational chemistry and high-performance computing. His work involves designing shared- and distributed-memory parallel CPU and GPU codes in C, C++, and CUDA, and conducting large-scale experiments on supercomputers.

SRP Project Title

Develop tensor algorithms in Python

Topical Areas

Applied Computer Science; Applied Mathematics; High Performance Computing; Open Source Software

Abstract

A wide range of applications involve multidimensional data as observations or solutions, and these data sets are often referred to as tensors. The storage cost of tensors grows exponentially with respect to the dimensionality, also known as "the curse of dimensionality", so researchers develop various data-sparse tensor formats for lower storage and faster computations. This project focuses on a special tensor format called the tensor-train (TT) format. We have developed some data-centric algorithms to factor a given tensor into TT representations. Our proposed algorithms can also be parallelized, and we can use them in analyzing large data sets, solving partial differential equations, and conducting statistical learning. In order to compare with state-of-the-art data-centric TT decomposition packages, we would like to implement our algorithms in Python, specifically with PyTorch and MPI4Py, using highly optimized linear algebra kernels and distributed memory parallelism. Then we can perform a thorough comparison between our proposed algorithms and the existing ones.

Desired Skills

Interests or background in numerical linear algebra and parallel computing. Desired skills: know how to code in Python, preferably PyTorch.

Lightning Talk Title

Tensor Computations in Python

Keywords

Numerical linear algebra; tensor computations; parallel algorithms; Python.