Structured Sparsity Specification Kit
This is a PyTorch implementation of the \(S^3\) framework.
Sparsekit provides a composable hierarchy for specifying and enforcing
structured sparsity patterns – from 2:4 and 4:8 to coupled multi-parameter
patterns. Pruning and masking write through directly to the
original nn.Parameter storage via torch.as_strided whenever possible to
avoid copies.
Sparse MMA pipeline: the full A matrix (32 x 16) with coupled 1:2 row sparsity (rows 8 apart, one pruned per pair) is compressed to the hardware A fragment (16 x 16), multiplied by the dense B fragment (8 x 16), producing a sparse output C (32 x 8) where pruned rows remain zero.
User Guide