Navaneeth Kunhi Purayil

At-the-Roofline Sparse Tensor Contractions on Vector Processors for Transformer Inference

Bowen Wang, Chi Zhang, Diyou Shen, Renzo Andri, Navaneeth Kunhi Purayil, Luca Benini

Fine-grained weight pruning and activation sparsification have emerged as effective approaches for reducing the compute and memory cost of inference for Transformer models. In the moderate-sparsity regime, Gustavson's dataflow provides a natural execution model for exploiting bot…Read more