Succeed to install pytorch by NVIDIA’s website(torch-1.1.0-cp36-cp36m-linux_aarch64.whl),
but my device only suits for cpu only version.NO GPU,NO CUDA.
import torch
Traceback (most recent call last):
File “”, line 1, in
File “/usr/local/lib/python3.6/dist-packages/torch/init.py”, line 79, in
from torch._C import *
ImportError: libcudart.so.10.0: cannot open shared object file: No such file or directory
So, how to import pytorch without cuda on aarch64 processor?Thank you.
No , python -version or pytorch -version is not necessary.
I use the pip3 at first , fail , wheel only applies to linux_x86_64/win_amd64.
Fortunately, NVIDIA provide aarch64’s whl for NVIDIA Jetson TX2,SO I want to turn cuda to no-cuda.
I almost give up this meaningless pursuits.
Ho sorry I missed the fact that you were on arm. We don’t provide official binaries.
But you can find in the issue tracker discussions about this here. And people providing their own binaries here.
Have you solved this problem
I have a khadas development board
and I encountered the same problem as you
Hello,
I face the same problem here.
I have compiled and installed the newest pytorch 1.6.0a0 on raspberry pi 4 4GB all by myself (I might change some choices in CMakeLists.txt) (Operating System: raspbian, armv7l)
but… I failed to compile it on ubuntu server on raspberry pi.
The building and installing procedure is just
python3 setup.py bdist_wheel
to get pytorch, and
python3 tools/build_libtorch.py
to build libtorch. (P.S. cmake+make is not recommended)
Now I give up building from myself. Perhaps it is the hardware that counts? (no SWAP partition in my OS?)
Use built wheel from others might be a good choice.
If you just want to use Libtorch? …
You can find TorchConfig.cmake, which is very essential. And that is the key to using libtorch.
The two clip above shows an pytorch installed on anaconda3 on my Microsoft Windows 10. Other operating systems (Linux, or some others) may be a little different.
For more instructions on using libtorch, see the libtorch’s turtorial.