[Solved] flash-attn reports error flash_attn_2_cuda.cpython-310-x86_64-linux-gnu.so: undefined symbol
posted on 2024-11-02 14:03 read(961) comment(0) like(10) collect(1)
Error message :
ImportError: /home/operationgpt/anaconda3/envs/lyj_py10_torch230/lib/python3.10/site-packages/flash_attn_2_cuda.cpython-310-x86_64-linux-gnu.so: undefined symbol: _ZN3c104cuda9SetDeviceEi
Solution: Reinstall flash attention
- Uninstall the existing flash-attn, enter
pip uninstall flash-attn, then enter y - Check your corresponding torch version, cuda version and python version
Check the torch version
pip show torch
The following results are returned, indicating that the torch version is 2.3.1
Check the cuda version
nvcc -V
The cuda version is V12.5.40
Check Python version
python --version
- Go to the flash attention official website to download the installation package. Note that you need to choose according to your torch version, cuda version (you can choose a version lower than your cuda version) and python version. Also select abiFALSE.
Right click, copy the link, and use wget + link to download the whl installation package in Linux :
wget https://github.com/Dao-AILab/flash-attention/releases/download/v2.6.3/flash_attn-2.6.3+cu123torch2.3cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
Finally, use pip install whl path to download flash-attn, and you're done!
Category of website: technical article > Blog
Author：Disheartened
link：http://www.pythonblackhole.com/blog/article/245758/5301600fea62ee1e2d9a/
source：python black hole net
Please indicate the source for any form of reprinting. If any infringement is discovered, it will be held legally responsible.
name:
Comment content: (supports up to 255 characters)
no articles