thu-ml / SageAttention

Quantized Attention that achieves speedups of 2.1-3.1x and 2.7-5.1x compared to FlashAttention2 and xformers, respectively, without lossing end-to-end metrics across various models.

Date Created 2024-10-03 (5 months ago)

Commits 85 (last one 3 days ago)

Stargazers 1,186 (35 this week)

Watchers 25 (0 this week)

Forks 78

License apache-2.0

Ranking

RepositoryStats indexes 630,443 repositories, of these thu-ml/SageAttention is ranked #45,512 (93rd percentile) for total stargazers, and #86,773 for total watchers. Github reports the primary language for this repository as Cuda, for repositories using this language it is ranked #26/394.

thu-ml/SageAttention is also tagged with popular topics, for these it's ranked: llm (#484/3538), cuda (#94/686), video-generation (#20/172)

Other Information

thu-ml/SageAttention has 2 open pull requests on Github, 10 pull requests have been merged over the lifetime of the repository.

Github issues are enabled, there are 33 open issues and 94 closed issues.

All Topics

llm cuda mlsys triton attention quantization video-generation efficient-attention inference-acceleration

Star History

Github stargazers over time

Watcher History

Github watchers over time, collection started in '23

Recent Commit History

85 commits on the default branch (main) since jan '22

Yearly Commits

Commits to the default branch (main) per year

Issue History

Languages

The primary language is Cuda but there's also others...

updated: 2025-03-24 @ 06:06am, id: 867007699 / R_kgDOM6180w