UST: The Universal Sparse Tensor

[aartbik]

The Universal Sparse Tensor (UST) simplifies sparse management by decoupling a tensor’s inherent sparsity from its memory storage representation. A tensor format Domain Specific Language (DSL) describes how a sparse tensor is represented. Type polymorphism on a small set of base operations defines the vast space of instances for these operations. The runtime inspects the chosen format to decide whether to dispatch to a heavily optimized library or use automatic sparse code generation when no existing solution is available.

We have already posted two blog posts on this subject:

• Establishing a Scalable Sparse Ecosystem with the Universal Sparse Tensor
• Simplify Sparse Deep Learning with Universal Sparse Tensor in nvmath-python

Furthermore, nvmath-python specifics can be found in the online documentation:

• Sparse Linear Algebra

However, since the UST is a relatively new concept, we plan to post a small series of articles on how to use the UST in nvmath-python. These articles will appear in this show-and-tell section of the forum, using the generic “UST: Topic” title.

Stay tuned! In our next installment, we’ll dive into diagonal formats and showcase how leveraging these specialized structures can unlock significant performance gains for your sparse workloads.