RCEpic
thu-ml/SageAttention
[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.
- Owner
- thu-ml (organization)
- Stars
- Stars: 3,991
- Forks
- Forks: 546
- Language
- Cuda
- License
- Apache-2.0
- Created
- 2024
- Topics
- attentioninference-accelerationllmquantizationcudatritonvideo-generationefficient-attentionmlsysllm-infra
- Tier
- RC (2,000 to 4,999 ★) What this tier means
More repos by thu-ml
All 27 repos by thu-ml- thu-ml/tianshouStars: 11,046PythonStable
- thu-ml/TurboDiffusionStars: 3,857PythonRC
- thu-ml/zhusuanStars: 2,218PythonRC
- thu-ml/RoboticsDiffusionTransformerStars: 1,816PythonBeta
- thu-ml/prolificdreamerStars: 1,563PythonBeta
- thu-ml/unidiffuserStars: 1,489PythonBeta
Nearby in Cuda
All Cuda repos- leoxiaobin/deep-high-resolution-net.pytorchStars: 4,476CudaRC
- Infatoshi/cuda-courseStars: 4,168CudaRC
- baidu-research/warp-ctcStars: 4,065CudaRC
- HazyResearch/ThunderKittensStars: 3,792CudaRC
- hujie-frank/SENetStars: 3,653CudaRC
- BBuf/how-to-optim-algorithm-in-cudaStars: 3,340CudaRC