FlashMLA
DeepSeekFlashMLA is DeepSeek's library of optimized attention kernels for serving its own models on NVIDIA GPUs and Huawei Ascend NPUs. It carries sparse (DeepSeek Sparse Attention) prefill and decoding kernels over FP8 or FP4 KV caches, dense multi-head attention kernels for the forward and backward passes, and a fused kernel for the small operations around attention. Support for Hopper GPUs and earlier DeepSeek models was dropped from the main line.
Openness
5 high confidence- license
- MIT(OSI)
- source
- public
- core features withheld
- no
MIT license body confirmed. The repository is public and unarchived and builds the whole product, and the README describes no paid tier, enterprise edition or license-gated build beside it, so source is public and the core ungated.
- https://api.github.com/repos/deepseek-ai/FlashMLA recorded 2026-09-02
Repo metadata for flashmla's canonical repository - private false, archived false, fork false.
- https://github.com/deepseek-ai/FlashMLA/blob/main/LICENSE recorded 2026-10-07
LICENSE body is the verbatim MIT text under a DeepSeek copyright line.
- https://github.com/deepseek-ai/FlashMLA/blob/main/README.md recorded 2026-10-07
README describes FlashMLA as "DeepSeek's library of optimized attention kernels, powering inference of the" DeepSeek-V4.1 model "on NVIDIA GPUs and Huawei Ascend NPUs", with sparse attention prefill and decoding kernels and a dense MHA prefill kernel; it names no paid tier, enterprise edition or license-gated build beside the published source.
Adoption
3 low confidenceThe `flash-mla` package on PyPI is a single unrelated placeholder release, not DeepSeek's distribution; FlashMLA's own README installs it from source instead. That leaves GitHub stars as the only adoption signal, which measures attention rather than confirmed use.
- https://api.github.com/repos/deepseek-ai/FlashMLA recorded 2026-09-02
Repo metadata - stargazers_count = 12,895 - for deepseek-ai/FlashMLA.
Capability
2 medium confidenceFlashMLA and FlashAttention are both attention-kernel libraries, but FlashMLA's kernels are purpose-built for DeepSeek's own multi-head latent and sparse attention rather than general-purpose use across architectures. That places it two steps below FlashAttention.
- https://github.com/deepseek-ai/FlashMLA/blob/main/README.md recorded 2026-10-07
README describes FlashMLA as "DeepSeek's library of optimized attention kernels, powering inference of the" DeepSeek-V4.1 model "on NVIDIA GPUs and Huawei Ascend NPUs", with sparse attention prefill and decoding kernels and a dense MHA prefill kernel; it names no paid tier, enterprise edition or license-gated build beside the published source.
Verified 2026-09-27