# deepseek-ai/FlashMLA

FlashMLA: Efficient Multi-head Latent Attention Kernels

Repository: https://github.com/deepseek-ai/FlashMLA
Canonical: https://ross.abutalabs.com/products/flashmla
Language: C++
License: MIT
License Family: permissive
Last push: 2026-07-28T06:18:26+00:00

## Health v2 (maintenance only)
Score: 62/100 (v2, computed 2026-09-02T17:46:02.011165+00:00)
- activity 94, release rhythm 35, longevity 39
- inputs: {"age_days": 558, "days_push": 36, "days_rel": null, "gap_med": null, "n_releases_24m": 0}
- flags: no_releases
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 12872, forks 1137 (observed 2026-08-28T04:10:59.961817+00:00)

## What it is
FlashMLA is DeepSeek's library of optimized CUDA attention kernels implementing Multi-head Latent Attention (MLA), including dense and token-level sparse attention for prefill and decoding stages with FP8 KV cache support. It powers DeepSeek-V3 and DeepSeek-V3.2 models and achieves up to 660 TFLOPS on H800 and 1460 TFLOPS on B200 GPUs.

## Use cases
- speed up MLA decoding for DeepSeek-style LLM inference
- run sparse attention with FP8 KV cache on Hopper GPUs
- benchmark attention kernel throughput on H800 or B200
- integrate optimized prefill attention kernels into an inference engine
- serve large language models with memory-efficient attention

## When to choose
- you are serving DeepSeek-V3/V3.2 or other MLA-based models on NVIDIA Hopper or Blackwell GPUs
- you need maximum attention throughput for prefill or decoding in a custom inference stack
- you want FP8 KV cache sparse decoding kernels

## When to avoid
- you use standard multi-head attention models without MLA support
- you run inference on non-NVIDIA hardware like AMD or Apple GPUs
- you need a turnkey inference server rather than low-level kernels

## Facets
- artifact type: library
- maturity: active
- function: machine-learning, llm-inference, gpu-computing, benchmarking
- domain: deep-learning, large-language-models, gpu-computing, performance
- platform: python, cpp
- tags: attention-kernels, cuda, mla, sparse-attention, fp8, flash-attention, hopper, blackwell, gpu, linux

## Member repositories
- deepseek-ai/FlashMLA (main) score 62

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:10:59.961817+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-29T17:13:43.354265+00:00, confidence not recorded.
  - readme: https://github.com/deepseek-ai/FlashMLA (fetched 2026-08-28T04:10:59.961817+00:00, sha 7f611d83d918)
- Data as of 2026-08-30T08:39:29.467469+00:00.