deepseek-ai/FlashMLA

FlashMLA: Efficient Multi-head Latent Attention Kernels

441 best points · 108 comments · 1 Hacker News post · 13,253 GitHub stars · View on GitHub

Hacker News discussions

• DeepSeek Open Source FlashMLA – MLA Decoding Kernel for Hopper GPUs (441 points, 108 comments, by helloericsf, 2025-02-24)