FlashAttention installation error: "CUDA 11.6 and above" requirement issue

[21X5122]

Hi,

I encountered an error while trying to install flash-attn (version 2.6.3) via pip. The error message indicates that FlashAttention requires CUDA 11.6 or above, but I am using PyTorch with CUDA 11.8 (torch.__version__ = 2.5.0+cu118). The installation fails during metadata generation, and I'm unsure how to proceed. Below is the full error output:

Collecting flash-attn
  Using cached flash_attn-2.6.3.tar.gz (2.6 MB)
  Preparing metadata (setup.py) ... error
  error: subprocess-exited-with-error

  × python setup.py egg_info did not run successfully.
  │ exit code: 1
  ╰─> [12 lines of output]
      fatal: not a git repository (or any of the parent directories): .git
      Traceback (most recent call last):
        File "<string>", line 2, in <module>
        File "<pip-setuptools-caller>", line 34, in <module>
        File "/tmp/pip-install-ur6gdyxq/flash-attn_8aaa3421ed454954b43cbcebfe9bb28e/setup.py", line 160, in <module>
          raise RuntimeError(
      RuntimeError: FlashAttention is only supported on CUDA 11.6 and above.  Note: make sure nvcc has a supported version by running nvcc -V.

      torch.__version__  = 2.5.0+cu118

      [end of output]

  note: This error originates from a subprocess, and is likely not a problem with pip.
error: metadata-generation-failed

× Encountered error while generating package metadata.
╰─> See above for output.

note: This is an issue with the package mentioned above, not pip.
hint: See above for details.

I have already checked my CUDA version using nvcc -V, and it shows CUDA 11.8 is installed.

Would you be able to provide any guidance or suggestions on how to resolve this issue? Is there anything else I should configure or check to ensure compatibility with flash-attn?

Thanks in advance for your help!

[skyprince999]

I am installing it with CUDA 12.2 and higher.