search

facebookresearch / hanabi_SAD

Simplified Action Decoder for Deep Multi-Agent Reinforcement Learning
Other
96 stars 35 forks source link

Compiled torch cannot find GPUs in Docker #6

Closed 51616 closed 4 years ago

51616 commented 4 years ago

I followed the instruction as in the README.md but the compiled torch cannot find GPUs devices.

The output after I ran sh tools/dev.sh Screenshot_2020-03-19_16-13-35

import torch
>>torch.cuda.is_available()
False
>>torch.backends.cudnn.enabled
True

If I use pip/conda torch, both will retrun True but the compiled C code cannot be imported similar to this. image

I tried different docker images (cuda 9.2, cuda 10.2) but all have the same behavior. All using gcc 7.5.0.

Update -- I tried on Ubuntu machine with gcc7.4.0 and cuda10.1 and also have the same problem :(

51616 commented 4 years ago

The problem is with the cuda-aware OpenMPI thing which is not installed by default on some linux system (I have no problem with linux mint 19.2). So installing cuda-aware openmpi solve this issue. :)