search

NVIDIA / MinkowskiEngine

Minkowski Engine is an auto-diff neural network library for high-dimensional sparse tensors
https://nvidia.github.io/MinkowskiEngine
Other
2.48k stars 367 forks source link

When building docker image as shown in the wiki only CPU support is built #379

Open weders opened 3 years ago

weders commented 3 years ago

Hi!

I would like to build a Docker container for the MinkowskiEngine. However, if I exactly follow the process described in the wiki, ME is only built for CPU support.

Afterwards, I tried to install it using --install-option="--force_cuda". Unfortunately, this led to another error

File "/usr/local/lib/python3.7/dist-packages/torch/utils/cpp_extension.py", line 1562, in _get_cuda_arch_flags arch_list[-1] += '+PTX' IndexError: list index out of range

Any idea what the problem could be?

chrischoy commented 3 years ago

It uses torch.cuda.is_available() to determine whether to use cuda compilation. Make sure the above command gets True.

VenSum007 commented 2 years ago

Hello @weders.. I have the same issue. Could you resolve it ?

weders commented 2 years ago

You have to adjust the setup.py file such that it does not use torch.cuda.is_available() as the GPU is not exposed to the docker during docker build .

The force_cuda flag should do the trick. You can find it here.

VenSum007 commented 2 years ago

Thanks for quick response @weders. I am not using docker. I cloned the repository and directly ran the setup.py using "python setup.py install --blas_include_dirs=${CONDA_PREFIX}/include --blas=openblas --force_cuda"

The problem still occurs though i used the force_cuda flag. I tried commenting the lines with torch.cuda.is_available() but found no luck.

weders commented 2 years ago

torch.cuda.is_available() returns True if you run it on your machine?

Something I also realized is that the CUDA versions have to perfectly match between PyTorch installation and CUDA Toolkit.

VenSum007 commented 2 years ago

torch.cuda.is_available() was returning False in my virtual environment. Matching the CUDA versions resolved the issue. @weders Thanks for your help :)