The problem with https://github.com/Dao-AILab/flash-attention/releases is that it shows 83 options and you have to be a DEEP Python/PyTorch/Linux expert to correctly pick the right one.

So ideally the README would include instructions (or a link to instructions) on how best to decide which wheel to use. Even better would be a little Python program you can run that tells you which one to use.

Here's a Python one-liner I got Claude to knock up which is almost certainly NOT correct but at least illustrates the concept:

python -c "
import sys, torch, platform;
py_ver = f'cp{sys.version_info.major}{sys.version_info.minor}';
cuda_ver = 'cu123' if torch.version.cuda and torch.version.cuda.startswith('12.3') else 'cu118';
abi = 'TRUE' if hasattr(sys, '_emscripten_info') or platform.libc_ver()[0] == 'glibc' else 'FALSE';
print(f'flash_attn-2.6.3+{cuda_ver}torch{torch.__version__.split(\"+\")[0]}cxx11abi{abi}-{py_ver}-{py_ver}-linux_x86_64.whl')
"

Outputs something like this:

flash_attn-2.6.3+cu118torch2.0.1cxx11abiFALSE-cp310-cp310-linux_x86_64.whl

Someone who knows what they're doing could write a version of this that works!

Huh, weird, it looks like setup.py is meant to automatically find them!

I wonder why that didn't work for me running a Google Colab notebook with a A100.

This PR is probably no good as it stands, but additional documentation to help people maximize the chance of using a pre-built wheel would definitely be useful.