在 Ubuntu18.04 中使用 CUDA 使用 Python 库 Keops 的运行时错误

Question

I am trying to run samples from the Python library: GeomLoss , which depends on CUDA, Pytorch and Keops in Ubuntu 18.04.3. I downloaded Python3.7 using Anaconda, and I am using CUDA 10.1. The gcc version is 7.4.0.

When I run samples from GeomLoss, the error message said:

RuntimeError: [KeOps] This KeOps shared object has been compiled without cuda support: try to set tagHostDevice to 0 or recompile the formula with a working version of cuda.

I cannot change tagHostDevice to 0 since this will disable GPU calculation according to their documentation. I checked CUDA and Pytorch installation and they was no error. But when I tried to run the installation checking code from KeOps:

import torch
import pykeops.torch as pktorch

x = torch.arange(1, 10, dtype=torch.float32).view(-1, 3)
y = torch.arange(3, 9, dtype=torch.float32).view(-1, 3)

my_conv = pktorch.Genred('SqNorm2(x-y)', ['x = Vi(3)', 'y = Vj(3)'])
print(my_conv(x, y)) 

I received error message:

error -- unsupported GNU version! gcc versions later than 6 are not supported! ^~~~~ CMake Error at keopslibKeOpstorch91c92bd508_generated_link_autodiff.cu.o.Release.cmake:219

I checked CUDA documentation, for Ubuntu 18.04.3, the native linux distribution support should be gcc-7.3.0. for x86_64. I used gcc --version to check default gcc in system and it is using gcc-7.4.0. I am not sure if this is the problem with using KeOps with CUDA and GPU. Also, I believe KeOps will not support gcc versions before 7. So I am really confused about what should I do to fix the problem right now.

I am wondering if anyone has experienced similar problems with GeomLoss and KeOps or other libraries. I am indeed grateful for any suggestions. Thanks!

Answer 1

I did the following steps and it worked for me:

First, by checking the dependencies in this link I noticed that nvcc compiler is not installed. By going to Nvidia Toolkit Installation Guide I did the following steps:

wget https://developer.download.nvidia.com/compute/cuda/11.2.2/local_installers/cuda_11.2.2_460.32.03_linux.run
sudo sh cuda_11.2.2_460.32.03_linux.run

Then I realized that nvcc command not working, so I did add them to the path using:

nano ~/.bashrc

# Add the following two lines: 
export PATH="$PATH:/usr/local/cuda/bin"
export LD_LIBRARY_PATH="$LD_LIBRARY_PATH:/usr/local/cuda/lib64"

In general, it is important to see the logs of PyKeops to check the process. So, you can always change the verbose and debug mode and see the details to check out what failed:

# Testing PyKeops installation
import pykeops

# Changing verbose and mode
pykeops.verbose = True
pykeops.build_type = 'Debug'

# Clean up the already compiled files
pykeops.clean_pykeops()

# Test Numpy integration
pykeops.test_numpy_bindings()