How to use system cuda/cudnn

1

I have a machine with cuda/cudnn compatible rocm device.

$ nvcc --version
HIPHSA: Author SUGON
HIP version: 5.4.23453
Cuda compilation tools, release 11.8, V11.8.89
clang version 15.0.0 (http://10.15.3.7/dcutoolkit/driverruntime/llvm-project.git 1be90618e508074abc746ab4963d7ad92710d6c5)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /public/software/compiler/dtk-23.10.1/llvm/bin

I then install pytorch 2.2 with cuda 11.8 by:

pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

But when I import torch, it can’t find cuda device:

$ python
Python 3.11.8 (main, Feb 26 2024, 21:39:34) [GCC 11.2.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.cuda.device_count()
0

I think the problem is pytorch use the cuda/cudnn runtime lib of its own. But I want it to use system cuda.

I have set CUDA_HOME and LD_LIBRARY_PATH. But it seems not work.

2

Which device are you using?
The binaries ship with their own CUDA runtime and your locally installed CUDA toolkit will be used if you build PyTorch from source.