@ICML

Researchers from NVIDIA have developed a new sparse data format and custom GPU kernels, called TwELL, that reshape unstructured sparsity in transformer language models to align with GPU architecture, …