MoonshotAI/FlashKDA

FlashKDA: high-performance Kimi Delta Attention kernels

Cuda1.2K stars118 forks+91 today

What it does

FlashKDA is a high-performance CUDA kernel for Kimi Delta Attention, optimized for SM90 and above GPUs with CUDA 12.9+. It integrates seamlessly with flash-linear-attention to enhance performance in AI applications.

Star history

Not enough history yet — 1 day(s) recorded. The daily snapshot builds this up.

Tracking

Last trending: 2026-08-02

Creator kit

Hook

🚀 Dive into the future of GPU-accelerated AI with FlashKDA, a cutting-edge CUDA kernel for Kimi Delta Attention. #AI #CUDA

Content angles

Create a tutorial on how to integrate FlashKDA into existing PyTorch projects.
Analyze performance benchmarks and discuss real-world applications in machine learning research.
Develop a comparative study of different GPU optimization techniques, including the benefits of using FlashKDA.

Who should care

AI researchers, developers working with CUDA and PyTorch, and anyone interested in high-performance computing for deep learning.