Flash Sparse Attention

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/HKUSTDial/flash-sparse-attention/blob/main/LICENSE recorded 2026-09-02

  LICENSE body is the verbatim three-clause BSD text under an HKUST(GZ) DIAL copyright line.

• https://api.github.com/repos/HKUSTDial/flash-sparse-attention recorded 2026-09-02

  Repo metadata for flash-sparse-attention's canonical repository - private false, archived false, fork false.

• https://github.com/HKUSTDial/flash-sparse-attention/blob/main/README.md recorded 2026-09-02

  README describes Flash-Sparse-Attention as a high-performance trainable sparse attention implementation supporting dense, sparse and gated attention, GQA/MQA, paged attention and fused low-precision quantization, with no paid tier, enterprise edition or license-gated build beside the published source.

64 PyPI downloads of `flash-sparse-attn` in the trailing 30 days, which lands in the <10K band of the software usage scale, level 1. The package is published from HKUSTDial/flash-sparse-attention and `pip install flash-sparse-attn` is the README's install path.

• https://pypistats.org/api/packages/flash-sparse-attn/recent recorded 2026-09-02

  last_month downloads = 64 for flash-sparse-attn

• https://pypi.org/pypi/flash-sparse-attn/json recorded 2026-09-02

  PyPI metadata for flash-sparse-attn 2.0.5 - project_urls Source = https://github.com/HKUSTDial/flash-sparse-attention, summary 'Flash Sparse Attention: Fast and Memory-Efficient Trainable Sparse Attention'.

Banded on the category feature matrix as one optimization family, applied broadly - the same shape as flash-attention, whose feature breadth (dense/sparse/gated variants, GQA/MQA, paged attention) it matches rather than narrowing to one model's attention variant. Placed two bands below the apache-tvm anchor.

• https://github.com/HKUSTDial/flash-sparse-attention/blob/main/README.md recorded 2026-09-02

  README describes Flash-Sparse-Attention as a high-performance trainable sparse attention implementation supporting dense, sparse and gated attention, GQA/MQA, paged attention and fused low-precision quantization, with no paid tier, enterprise edition or license-gated build beside the published source.