4

I would like to go to CUDA (cudatoolkit) version compatible with Nvidie-430 driver, i.e., 10.0.130 as recommended by the Nvidias site.

Based on this answer I did,

conda install -c pytorch cudatoolkit=10.0.130

And then I get this error (pastebin link). (very-short version below):

(fastaiclean) eghx@eghx-nitro:~$ conda install -c pytorch cudatoolkit=10.0.130
...  
Solving environment: failed with initial frozen solve. Retrying with flexible solve.
Solving environment: | 
Found conflicts! Looking for incompatible packages.
This can take several minutes.  Press CTRL-C to abort.
failed   
UnsatisfiableError: The following specifications were found to be incompatible with each other:

Output in format: Requested package -> Available versions

Package _libgcc_mutex conflicts for:
pyzmq -> libgcc-ng[version='>=7.3.0'] -> _libgcc_mutex=[build=main]
libgcc-ng -> _libgcc_mutex=[build=main]
lcms2 -> libgcc-ng[version='>=7.3.0'] -> _libgcc_mutex=[build=main]
...  
The following specifications were found to be incompatible with your system:

  - feature:/linux-64::__cuda==10.1=0
  - feature:|@/linux-64::__cuda==10.1=0

Your installed version is: 10.1

Why am I getting conflicts? Why does it say 10.1 when cuda toolkit is 10.2.89 (conda list)? how to handle conflicts? What can I do with this error? The conflicts are so huge, I don't know where to start.

Other

  • Nvidia driver 430

  • current cudatoolkit: 10.2.89

talonmies
  • 70,661
  • 34
  • 192
  • 269
agent18
  • 2,109
  • 4
  • 20
  • 34
  • I remember installed Pytorch via pip with incorrect Cuda version (10.1 vs 11) and it still worked. IIRC `pip install torch` also install `libcudarc.so` and other things with it too. If it works then don't fix it – Natthaphon Hongcharoen Sep 27 '20 at 16:05
  • pytorch and cuda might be compatible. But the thing is it is the nvidia driver (version-430) which doesn't work with 10.1 and above ([Nvidia website](https://docs.nvidia.com/deploy/cuda-compatibility/index.html#binary-compatibility)). Can you let me know a list of what you think I should install? What does "other things with it" mean? I am a noob. Should I also uninstall pytorch and cudatoolkit? or? – agent18 Sep 27 '20 at 17:15
  • Nah, I just `pip install torch` back then and it installed cuda 10.1 which is not suppose to be usable on that PC. You should try `import torch; torch.zeros(1).cuda()` and see if it error or not. If not then it works – Natthaphon Hongcharoen Sep 27 '20 at 17:32
  • `AssertionError: The NVIDIA driver on your system is too old (found version 10010).` I think this tells that Nvidia driver and cuda are not compatible. But, I can't increase my nvidia driver version as for ubuntu 16 it is only 430 (which is the latest). – agent18 Sep 27 '20 at 17:35
  • So I tried to go to 10.0.130 – agent18 Sep 27 '20 at 17:35
  • Does `pip install torch==1.5.1+cu92 torchvision==0.6.1+cu92 -f https://download.pytorch.org/whl/torch_stable.html ` work? That's all I can help though – Natthaphon Hongcharoen Sep 27 '20 at 17:40
  • @natthaphonhongcharoen thanks for the help. I didn't try it... I wanted to stick with conda and see if I got a solution first and did find something. – agent18 Sep 30 '20 at 21:34

1 Answers1

1

Check current version with

torch.version.cuda

I had 10.2. But I need 10.1 according to: table 1 here and my 430 NVIDIA driver installed.

Uninstall and Install

conda remove pytorch torchvision cudatoolkit

conda install pytorch==1.6.0 torchvision==0.7.0 cudatoolkit=10.1.168 -c pytorch

Say yes to everything for the above commands.

agent18
  • 2,109
  • 4
  • 20
  • 34