First, I have to thank MinChoi0129 for the information shared in this issue.

As a beginner trying to build the flash-attn package for the first time, I ran into the same failures as many other people. If your Python version, PyTorch version, and CUDA version are not chosen carefully, the install falls back to build-from-source mode. That mode not only consumes a huge amount of memory (~40 GB) and a lot of time (30 minutes to 1 hour), but also often ends in a failed build.

So here is one version combination that worked for me.

Hardware:

• GPU: RTX 5080

System:

• WSL 2 (Ubuntu 22.04.5)

Software:

• Python = 3.10
• CUDA = 12.8
• torch = 2.7.1
  (pip install torch==2.7.1 torchvision==0.22.1 torchaudio==2.7.1 --index-url https://download.pytorch.org/whl/cu128)

Flash Attention version:

• flash-attn = 2.8.3

Installation command:

pip install flash-attn torch==2.7.1 --no-build-isolation

With this combination, memory usage was almost unchanged and the installation succeeded in about one minute.

贡献者: Junyuan He