pip installation doesn't include cuda version

[torch-points3d]

pip installation doesn't include cuda version

[nicolas-chaulet]

Hi,
This is probably because it cannot find CUDA. What does the following python code give you?

from torch.utils.cpp_extension import  CUDA_HOME
print(CUDA_HOME)

It should return the path to your CUDA installation.

[pvnieo]

Hi,
It gives the following: /usr.

[nicolas-chaulet]

Ok, something is wrong then. Could you add you cuda bin directory to the PATH environment variable? It should be something like

export PATH="/usr/local/cuda-10.2/bin:$PATH"

and run the command again?

[zeqinLi]

Hi,I have the same problem.When i run the code:

from torch.utils.cpp_extension import  CUDA_HOME
print(CUDA_HOME)

The code return the result:

public/software/cuda/10.1

But when i import torch_points_kernels.points_cuda,the following error occurred:

ModuleNotFoundError: No module named 'torch_points_kernels.points_cuda'

In the path(/public/software/anaconda3-5.2.0/envs/pytorch/lib/python3.6/site-packages/torch_points _kernel) there are only the following files：

The system I use is centoros. How can i solve this problem?

[nicolas-chaulet]

@pvnieo Are you using conda to create the virtual env and to install torch-points-kernels? If yes then that is probably the issue. I haven't managed to fix conda installs. I would recommend using plain virtual envs until this is fixed (contributions welcome, I am not very familiar with conda!)

python -m venv tp3denv
source tp3denv/bin/activate
pip install torch
pip install torch-points-kernels

This should work just fine.

[pvnieo]

@nicolas-chaulet No I'm not using conda or virtual env, I installed the library using pip.
Also, I tried to find the bin for cuda but I didn't find it. The path you provided is empty.

[nicolas-chaulet]

What does nvidia-smi give you? And

import torch
print(torch.cuda.is_available())

It looks like it cannot find cuda properly...

[pvnieo]

torch.cuda.is_available() gives True.
nvidia-smi gives the details about my gpu and CUDA version:

+-----------------------------------------------------------------------------+
| NVIDIA-SMI 440.64       Driver Version: 440.64       CUDA Version: 10.2     |
|-------------------------------+----------------------+----------------------+
| GPU  Name        Persistence-M| Bus-Id        Disp.A | Volatile Uncorr. ECC |
| Fan  Temp  Perf  Pwr:Usage/Cap|         Memory-Usage | GPU-Util  Compute M. |
|===============================+======================+======================|
|   0  GeForce GTX 106...  Off  | 00000000:01:00.0 Off |                  N/A |
| N/A   58C    P5     8W /  N/A |    836MiB /  6078MiB |      1%      Default |
+-------------------------------+----------------------+----------------------+

[nicolas-chaulet]

Alright, that's weird. Could you post the logs of

pip uninstall torch-points-kernels
pip install torch-points-kernels --no-cache

please? Just trying to udnerstand what's going on...

[nicolas-chaulet]

I have added some troubleshooting steps in teh README, would you mind trying and letting me know if that helps?|
https://github.com/nicolas-chaulet/torch-points-kernels#troubleshooting

[pvnieo]

Hi, I have reinstalled torch kernel points in an another server, but this time I didn't got the error message. I'll let you know if I can reproduce it!

[pvnieo]

Hi @nicolas-chaulet ,
I got the same issue again, here's the output of:

from torch.utils.cpp_extension import  CUDA_HOME
print(CUDA_HOME)

/usr/local/cuda

[nicolas-chaulet]

I am pretty certain it is related to PATH variable, could you run through the trouble shooting steps here?
https://github.com/nicolas-chaulet/torch-points3d#cannot-compile-certain-cuda-kernels-or-seg-faults-while-running-the-tests

[yuminhhuang]

I am also with this problem [ ModuleNotFoundError: No module named 'torch_points_kernels.points_cuda' ] in one of my server but no problem in another server.
In my case, in one sever with points_cpu.so and points_cuda.so in lib/python3.9/site-packages/torch_points_kernels/, it will be OK. But in the other server only with points_cpu.so and without points_cuda.so will cause the ERROR. It is useless if I just copy the points_cuda.so to where it miss.

[yuminhhuang]

I solve my problem because the setup.py will check whether there is GPU in env. So I need to allocate GPU first in cluster and then build this project use setup.py. Then in my case the built result will not auto generate to lib/python3.9/site-packages/torch_points_kernels/, so I copy torch_points_kernels/ directory to there manually. Finally both points_cpu.so and points_cuda.so prepared and it works.