FlashAttention 2.8.3 — cu130 / Torch 2.11 / CPython 3.12

FlashAttention 2.8.3 binary for the pinned RobustOps environment.

This is a repack of the existing verl_r installation built on 2026-09-17, not a fresh source build or an upstream official wheel. All 104 packaged original files were checked against the installed distribution's RECORD hashes before repacking. Upstream license and author files are included.

Compatibility:

• Linux x86_64, CPython 3.12.
• PyTorch 2.11.0+cu130, CUDA runtime 13.0, C++11 ABI enabled.
• Compiled GPU targets: sm_80 and sm_120. No sm_90 or sm_100 kernels.
• The extension requires GLIBC symbols through 2.32 and GLIBCXX symbols through 3.4.21; dependencies can impose additional requirements. Build/validation host uses glibc 2.35.
• This is a linux_x86_64 wheel, not an audited manylinux wheel. Other distributions and GPUs have not been validated by this release.

Repack changes: a descriptive local package version, torch==2.11.0 metadata, removal of bytecode/installer state, regeneration of RECORD, and removal of the extension's absolute verl_r RPATH from the published copy. Original verl_r files were not modified.

Validation details and exact artifact hash are attached as validation logs, build-info.json, and SHA256SUMS. Use SHA256SUMS to verify the downloaded wheel before installation. The validation script is also attached.

Install into an environment that already contains the matching PyTorch/CUDA stack:

python -m pip install --no-deps \
  'https://github.com/haohuiliang/RobustOps/releases/download/flash-attn-2.8.3-cu130-torch2.11-cp312/flash_attn-2.8.3%2Bcu130.torch2.11.abi1.sm80sm120-cp312-cp312-linux_x86_64.whl#sha256=12154159f06ec92f6bd6c120b915a3f839b4ddccf910319c8ba63e869b4a8d82'

The URL includes the wheel's SHA256 digest for pip verification. Alternatively, download the wheel and SHA256SUMS from the release assets, run sha256sum -c SHA256SUMS, and install with python -m pip install --no-deps <wheel-file>.