| license: bsd-3-clause | |
| tags: | |
| - python-wheels | |
| - flash-attention | |
| # sie-wheels | |
| Prebuilt Python wheels that the [Superlinked Inference Engine](https://github.com/superlinked/sie) build pins by exact URL and sha256. They are unmodified copies of upstream builds, kept here so that CI and image builds do not depend on GitHub release downloads. | |
| Files are never changed after upload. Consumers reference a repository revision (`/resolve/<commit>/<file>`), so a later upload cannot change what an existing lockfile installs, and the lockfile records each file's sha256. | |
| ## Files | |
| | File | Upstream | sha256 | | |
| |---|---|---| | |
| | `flash_attn-2.7.4+cu129torch2.9-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl` | [mjun0812/flash-attention-prebuild-wheels v0.7.11](https://github.com/mjun0812/flash-attention-prebuild-wheels/releases/tag/v0.7.11), published 2026-01-17: [flash-attention](https://github.com/Dao-AILab/flash-attention) 2.7.4 built for torch 2.9 and CUDA 12.9, CPython 3.12, manylinux x86_64 | `0a8f6219e61f8ff923cfd2cdad105d95a66106a6c9e754486d778c3a25e7707f` | | |
| ## License | |
| flash-attention is distributed under the BSD 3-Clause License, included as `LICENSE`. The upstream wheel build scripts are BSD 3-Clause as well. | |
| ## Adding a wheel | |
| 1. Download the upstream file and record its sha256. | |
| 2. Upload it unchanged, together with an updated table above, in one commit. | |
| 3. Reference `https://huggingface.co/superlinked/sie-wheels/resolve/<commit>/<file>` and let the lockfile record the hash. | |