stanford-stk 0.0.3

Sparse Toolkit

:abacus: Sparse Toolkit

A light-weight PyTorch library for block-sparse matrices and block-sparse matrix multiplication.

STK is built around a core sparse matrix class (stk.Matrix), which uses a hybrid blocked-CSR-COO sparse matrix encoding to enable efficient matrix products with sparse inputs and outputs in transposed or non-transposed order. The library supports the following operations with block_size=128:

op: transpose or non-transpose

[Sparse Matrix Multiplication]
stk.ops.dsd: dense = op(sparse) x op(dense)
stk.ops.dds: dense = op(dense) x op(sparse)
stk.ops.sdd: sparse = op(dense) x op(dense)

stk.ops.ssd: sparse = op(sparse) x op(dense) [Experimental]
stk.ops.sds: sparse = op(dense) x op(sparse) [Experimental]
stk.ops.dss: dense = op(sparse) x op(sparse) [Experimental]

[Sparse Matrix Conversion]
stk.ops.to_sparse: torch.Tensor => stk.Matrix
stk.ops.to_dense: stk.Matrix => torch.Tensor

[Sparse Matrix Generation]
stk.random.dense_mask: Create a random, block-sparse dense matrix.
stk.random.mask: Create a random, block-sparse sparse matrix.

STK is designed for applications where the sparse matrices change rapidly. This is complementary to libraries like triton-blocksparse, which assume that sparse matrix topologies do not change between invocations.

:rocket: Performance

Block-sparse matrix multiplication with STK is able to match the performance of cuBLAS on a range of problems. On these benchmarks from MegaBlocks dMoE models, STK realizes 98.6% of cuBLAS throughput with 128x128 blocks on average.

Hardware: A100-SXM4-80GB
Software: CUDA 11.5, CUTLASS 2.5

:building_construction: Installation

pip install stanford-stk

:writing_hand: Citation

@article{megablocks-arxiv,
  author    = {Trevor Gale and Deepak Narayanan and Cliff Young and Matei Zaharia},
  title     = {MegaBlocks: Efficient Sparse Training with Mixture-of-Experts},
  journal   = {CoRR},
  volume    = {abs/2211.15841},
  year      = {2022},
}