[Wheel] Pre-built flash-attn 2.8.3 for CUDA 12 + PyTorch 2.11 (Python 3.10-3.13)

[lesj0610]

Since there is no official pre-built wheel for PyTorch 2.11 (related: #2299), I built Docker-based wheels and uploaded them to a fork release.

Release page: v2.8.3-cu12-torch2.11

Download

• Python 3.10: flash_attn-2.8.3+cu12torch2.11cxx11abiTRUE-cp310-cp310-linux_x86_64.whl
• Python 3.11: flash_attn-2.8.3+cu12torch2.11cxx11abiTRUE-cp311-cp311-linux_x86_64.whl
• Python 3.12: flash_attn-2.8.3+cu12torch2.11cxx11abiTRUE-cp312-cp312-linux_x86_64.whl
• Python 3.13: flash_attn-2.8.3+cu12torch2.11cxx11abiTRUE-cp313-cp313-linux_x86_64.whl

Example:

pip install "https://github.com/lesj0610/flash-attention/releases/download/v2.8.3-cu12-torch2.11/flash_attn-2.8.3%2Bcu12torch2.11cxx11abiTRUE-cp312-cp312-linux_x86_64.whl"

Build environment

flash-attention	v2.8.3
CUDA	12.8
PyTorch	2.11 (stable)
Python	3.10, 3.11, 3.12, 3.13
CXX11 ABI	TRUE
OS	Ubuntu 24.04

Notes

• Community build only.
• Python 3.9 is not included because upstream torch 2.11.0+cu128 does not publish Linux x86_64 cp39 wheels.

[HelloWorldLTY]

Thank you! Can we use torch 2.12 + cuda12.9 for this wheel?

[lesj0610]

Thank you! Can we use torch 2.12 + cuda12.9 for this wheel?

only ensures that it works in the version described

[TianHongZXY]

Could you provide wheels built for aarch64? Thanks in advance!

[kaka-Cao]

i use py3.11, which vllm version shuold i choose? and why no py3.11 with torch2.10? thanks!

[remixer-dec]

I'm getting ImportError: site-packages/flash_attn_2_cuda.cpython-311-x86_64-linux-gnu.so: undefined symbol: _ZN3c104cuda29c10_cuda_check_implementationEiPKcS2_jb on CUDA 12.8

UPD: oh, this build is for Pytorch 2.11 and I need 2.9 and nobody has built it in this config... #2212 what a time to be alive

[Apprisco]

Why not cuda 13 btw?

[EthanJi29]

i love you bro