Infrastructure / Compilers & Model Optimization

Block Sparse Attention

Block Sparse Attention is a library of sparse attention kernels from MIT HAN Lab supporting streaming attention at token and block granularity plus block-sparse attention, modified from FlashAttention to reduce the compute and memory cost of long-context LLM inference.

Verified 2026-09-02 via the GitHub API and the LICENSE body.

Openness

5 high confidence

5.0

license: BSD-3-Clause(OSI)
source: public
core-gated: ungated

BSD-3-Clause license body confirmed. The repository is public and unarchived and builds the whole product, and the README describes no paid tier, enterprise edition or license-gated build beside it, so source is public and the core ungated.

https://github.com/mit-han-lab/Block-Sparse-Attention/blob/main/LICENSE recorded 2026-09-02
LICENSE body is the verbatim three-clause BSD text, inherited from its FlashAttention base.
https://api.github.com/repos/mit-han-lab/Block-Sparse-Attention recorded 2026-09-02
Repo metadata for block-sparse-attention's canonical repository - private false, archived false, fork false.
https://github.com/mit-han-lab/Block-Sparse-Attention/blob/main/README.md recorded 2026-09-02
README describes Block Sparse Attention as a library of sparse attention kernels supporting streaming and block-sparse patterns, initially modified from FlashAttention 2.4.2, with no paid tier, enterprise edition or license-gated build beside the published source.

Adoption

1 low confidence

1.0

549 GitHub stars, which lands in the <1K stars band of the stars scale, level 1. Block-Sparse-Attention has no registry package and installs from source, so no download count exists and the stars scale applies.

https://api.github.com/repos/mit-han-lab/Block-Sparse-Attention recorded 2026-09-02
Repo metadata - stargazers_count = 549 - for mit-han-lab/Block-Sparse-Attention.

Capability

2 medium confidence

2.0

Banded on the category feature matrix as a narrow kernel set: a small number of sparse-pattern variants built on top of FlashAttention rather than the general attention coverage flash-attention itself supplies. Placed two bands below the tensorrt rung.

https://github.com/mit-han-lab/Block-Sparse-Attention/blob/main/README.md recorded 2026-09-02
README describes Block Sparse Attention as a library of sparse attention kernels supporting streaming and block-sparse patterns, initially modified from FlashAttention 2.4.2, with no paid tier, enterprise edition or license-gated build beside the published source.

Verified 2026-09-02