cuLA: Rewriting Linear Attention in CUDA
AI summary
cuLA is a high-performance CUDA kernel library for linear attention variants such as GLA and KDA. Using hand-optimized CuTe DSL and CUTLASS C++ implementations, it is deeply adapted to NVIDIA Hopper and Blackwell GPU architectures, aiming to turn complex algorithms into directly callable high-performance operators. Its interface is consistent with flash-linear-attention, so developers can integrate at low cost by changing one import line. Performance tests show its kernels achieve up to 1.52x average speedup over Triton baseline implementations on Blackwell and Hopper GPUs. The project is open-source and plans future integration into FLA's unified scheduling mechanism.
Key Takeaways
- cuLA is a high-performance CUDA kernel library for linear attention variants such as GLA and KDA.
- Using hand-optimized CuTe DSL and CUTLASS C++ implementations,it is deeply adapted to NVIDIA…
- Its interface is consistent with flash-linear-attention,so developers can integrate at low cost by…
- Performance tests show its kernels achieve up to 1.
- 52x average speedup over Triton baseline implementations on Blackwell and Hopper GPUs.
Quick Facts
- Published
- Updated
- Source
- 蚂蚁百灵:Developer Blog(网页)
- Category
- Products & Apps
- AI score
- 66
蚂蚁百灵把线性注意力的 CUDA 内核做成了开箱即用的库,性能超过 Triton 基线,对于做长上下文推理优化的团队挺实用,但受众面比较窄。
Related entities
Source & copyright
This article is based on curated titles and AI summaries from the AIHOT public API, with attribution and links back. Original copyright remains with the original publishers and authors; this site does not mirror full texts.
How to Cite
When citing this article, keep the title, date, and permalink, and credit Orange News as the source.
Orange News: cuLA: Rewriting Linear Attention in CUDA. April 8, 2026. https://new.nat6.net/en/news/cmoecv1p6003qsl19n5usdt4x/