deepseek-ai/flashmla is DeepSeek's optimized attention kernel library. According to this repository's README description, it supports DeepSeek-V4.1 inference on NVIDIA GPUs and Huawei Ascend NPUs. The provided FACTS contain no latest release version, so it cannot be introduced as a version update. What can currently be confirmed is the repository description and the features recorded in the README.
This project provides kernels for MLA and token-level sparse attention computation. Sparse attention is an approach that selects the token relationships that are needed instead of computing the relationships of all tokens. The README lists sparse attention prefill and decoding kernels that use FP8 and FP4 KV caches, fused kernels, and dense attention prefill and backward kernels. So the supported scope covers not only sparse attention but also some computation stages of dense attention. The implementation language is C++ and the license is MIT. The GitHub star count is 13,180. This is the scope as stated in the repository's public description.
The provided FACTS contain no performance measurements or comparison baselines. Therefore, while the supported platforms and kernel types can be confirmed, how much the actual processing speed improves cannot be judged from this material alone. The latest release linked to the repository is also not presented, making it difficult to assert the feature description as new changes in a specific version.