Pytorch is broken with cuda 11-6

Torch still trys to find libcupti.so.11.5 [1] and fails to find it.

Perhaps a rebuild on pytorch is necessary.

A similar thread in the Archlinux forum [2] suggests installing cuda-tools, but the issue still perisists on my end (possibly due to versions or differences between manjaro and arch).


[1]
Python 3.10.2 (main, Jan 15 2022, 19:56:27) [GCC 11.1.0] on linux
Type “help”, “copyright”, “credits” or “license” for more information.

import torch
Traceback (most recent call last):
File “”, line 1, in
File “/usr/lib/python3.10/site-packages/torch/init.py”, line 197, in
from torch._C import * # noqa: F403
ImportError: libcupti.so.11.5: cannot open shared object file: No such file or directory


[2] https_bbs.archlinux.org_viewtopic.php?pid=2022049
(They say I cannot include links so…)

[Edit log 1]
Reformat.
cuda-tookit → cuda-tools
Sry for the typo. I installed the cuda-tools in the community repo, not the cuda-toolkit from Nvidia.

Currently, I downgrade CUDA to 11.5.1 to solve the problem.
Plus, the dataloader bug [1,2] can be solved via [2].
The file need to modify is:
/usr/lib/python3.10/site-packages/torch/utils/data/_typing.py
and you can use my modification [3] if you have trouble following [2].

[1] https:_bugs.archlinux.org/task/73187
[2] https:_github.com/pytorch/pytorch/issues/69911#issuecomment-1007571327
[3] https:_pastebin.com/BEpH7wYj
Always review before using/running code from the internet~
They still do not allow me to include links, duh.

There’s an open Arch bug report: FS#73668 - python-pytorch-cuda needs to be recompiled with cuda 11.6