FlashAttention-4 (CuTeDSL)

FlashAttention-4 is a CuTeDSL-based implementation of FlashAttention for Hopper and Blackwell GPUs.

Installation

pip install flash-attn-4

If you're on CUDA 13, install with the cu13 extra for best performance:

pip install "flash-attn-4[cu13]"

Usage

from flash_attn.cute import flash_attn_func, flash_attn_varlen_func

out = flash_attn_func(q, k, v, causal=True)

Development

git clone https://github.com/Dao-AILab/flash-attention.git
cd flash-attention
pip install -e "flash_attn/cute[dev]"       # CUDA 12.x
pip install -e "flash_attn/cute[dev,cu13]"  # CUDA 13.x (e.g. B200)
pytest tests/cute/

Metadata

Release files for tokenspeed-fa4 4.0.0.post20260923

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Built distribution (wheel)

Table of built distributions (wheels) for tokenspeed-fa4 4.0.0.post20260923
File	Interpreter	ABI	Platform	Reset
tokenspeed_fa4-4.0.0.post20260923-py3-none-any.whl 402.9 kB Sep 23, 2026	Python 3	none	any	Details

Release files / tokenspeed_fa4-4.0.0.post20260923-py3-none-any.whl

Download URL	tokenspeed_fa4-4.0.0.post20260923-py3-none-any.whl
Size	402.9 kB
Tags	Python 3
SHA-256 checksum How to use checksums	`cf429aa1d35ee5bacbeb671d6493e95a63c0f38783bd631281e4229a60bb791c`
BLAKE2b-256 checksum How to use checksums	`f8e48e4efaca38b31584eaa94974c63039726dbbde3a375bcc0f4103cafba87b`
Upload date	Sep 23, 2026
Uploaded using Trusted Publishing? What is trusted publishing?	Yes
Uploaded via	`twine/7.0.0 CPython/3.13.14`