Infrastructure / Compilers & Model Optimization

SageAttention

SageAttention is a quantized attention implementation that replaces the attention operator with 8-bit and INT4 arithmetic, using outlier smoothing and per-thread quantization to hold accuracy. It is applied as a drop-in substitution in language, image and video models, and the repository carries the SageAttention, SageAttention2 and SageAttention2++ implementations described in the papers.

Verified 2026-08-18 via the GitHub API and the LICENSE body.

Openness

5 high confidence

5.0

license: Apache-2.0(OSI)
source: public
core-gated: ungated

Apache-2.0 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/thu-ml/SageAttention/blob/main/LICENSE recorded 2026-08-18
LICENSE file is the verbatim Apache-2.0 text
https://api.github.com/repos/thu-ml/SageAttention recorded 2026-08-18
Repo metadata - license spdx_id Apache-2.0, private false, archived false, default branch main - for thu-ml/SageAttention.
https://github.com/thu-ml/SageAttention/blob/main/README.md recorded 2026-08-18
README describes the official implementation of SageAttention, SageAttention2 and SageAttention2++, quantized 8-bit and INT4 attention applied as a plug-and-play replacement across language, image and video models, with no paid tier, enterprise edition or license-key-gated build beside it.

Adoption

3 high confidence

3.0

209,877 PyPI downloads of `sageattention` in the trailing 30 days, which lands in the 100K-1M band of the software usage scale, level 3.

https://pypistats.org/api/packages/sageattention/recent recorded 2026-08-18
last_month downloads = 209,877 for sageattention

Capability

2 medium confidence

2.0

Banded on the category feature matrix as narrow kernel set or single-pass utility. Placed two bands below the tensorrt anchor, on a matrix that bands on how much of the model-to-hardware transformation pipeline a product performs, over how many inputs and targets.

https://github.com/thu-ml/SageAttention/blob/main/README.md recorded 2026-08-18
README still documents the official implementation of SageAttention, SageAttention2 and SageAttention2++, quantized 8-bit and INT4 attention applied as a plug-and-play replacement across language, image and video models.

Verified 2026-08-18