Flash Attention index

PEP 503 indexes for flash-attention prebuilt wheels

These indexes provide prebuilt wheels for flash-attn to speed up installation and avoid compilation issues. The wheels are prebuilt and provided by @mjun0812 (and a few others by me)

Usage

You can use these indexes to easily install flash-attn with tools that support PEP 503 indexes, such as uv, poetry, or standard pip.

uv (in pyproject.toml)

[tool.uv.sources]
flash-attn = [{index = "flash-attn-cu130", marker = "sys_platform == 'linux' or sys_platform == 'win32'" }]

[[tool.uv.index]]
name = "flash-attn-cu130"
url = "https://codcordance.github.io/flash-attn-index/cu130/"
explicit = true

poetry (in pyproject.toml)

[[tool.poetry.source]]
name = "flash-attn-cu130"
url = "https://codcordance.github.io/flash-attn-index/cu130/"
priority = "explicit"

[tool.poetry.dependencies]
flash-attn = { version = "*", source = "flash-attn-cu130", markers = "sys_platform == 'linux' or sys_platform == 'win32'" }

pip

pip install flash-attn --extra-index-url https://codcordance.github.io/flash-attn-index/cu130/

(Note: Replace cu130 with cu126, cu128 or cu131 in the URLs and index names depending on your CUDA version requirements.)