STen: An Interface for Efficient Sparsity in PyTorch


METADATA ONLY
Loading...

Date

2022-07

Publication Type

Conference Paper

ETH Bibliography

yes

Citations

Altmetric
METADATA ONLY

Data

Rights / License

Abstract

As deep learning models grow, sparsity is becoming an increasingly critical component of deep neural networks, enabling improved performance and reduced storage. However, existing frameworks offer poor support for sparsity. They primarily focus on sparse tensors in classical formats such as COO and CSR, which are not well suited to the sparsity regimes typical of deep learning, and neglect the broader sparsification pipeline necessary for using sparse models. To address this, we propose a new sparsity interface for PyTorch, STen, that incorporates sparsity layouts for tensors (including parameters and transients, e.g., activations), sparsity-aware operators, and sparsifiers, which define how a tensor is sparsified, and supports virtually all sparsification methods. STen can enable better sparse performance and simplify building sparse models, helping to make sparsity easily accessible.

Publication status

published

Editor

Book title

Journal / series

Volume

Pages / Article No.

Publisher

ETH Zurich, Scalable Parallel Computing Laboratory

Event

2nd Sparsity in Neural Networks Workshop (SNN 2022)

Edition / version

Methods

Software

Geographic location

Date collected

Date created

Subject

Organisational unit

03950 - Hoefler, Torsten / Hoefler, Torsten check_circle

Notes

Poster presentation

Funding

Related publications and datasets