FlashMLA is DeepSeek's library of optimized attention kernels, powering the DeepSeek-V3 and DeepSeek-V3.2-Exp models. This repository contains the following implementations:
功能特性
- Token-level sparse attention for the prefill stage
- Token-level sparse attention for the decoding stage, with FP8 KV cache
- Dense attention for the prefill stage
- Dense attention for the decoding stage
- 2025.08.01 Kernels for MHA on SM100: Thanks to NVIDIA's PR for MHA forward / backward kernels on SM100!
- 2025.04.22 Deep-Dive Blog: We'd love to share the technical details behind the new FlashMLA kernel! Check out our deep-dive write-up here.