thu-ml/
[ICLR2025, ICML2025, NeurIPS2025 Spotlight] Quantized Attention achieves speedup of 2-5x compared to FlashAttention, without losing end-to-end metrics across language, image, and video models.
by TSAIL group Cuda Apache-2.0 licence 554 forks Created Oct 2024 View on GitHub
- Stars
- 4,007
- 11% of thu-ml's stars
- Rank
- #16,575
- of 851,979 repos, top 2%
Among 851,979 repos
4k–6.3k stars: 6,646 repos, including SageAttention
top 50%top 10%top 1%top 0.1% 
Stars along the bottom, and repos in each band on a log scale
Quick stats
- Next rank
- +1 star to reach #16,568
- Next milestone
- +3,208 stars to reach the top 1%
- Popular
- One of 66,555 repos with 1k+ stars
- Yearly average
- 1,897 stars a year since Oct 2024