Cuda error while initializing SparseTensor

[victoryc]

I am getting the following error while initializing a SparseTensor:

terminate called after throwing an instance of 'thrust::system::system_error'
  what():  CUDA free failed: cudaErrorIllegalAddress: an illegal memory access was encountered

Below is a small bit of code that reproduces this error [This is derived from the sparse_tensor_basic.py file in MinkowskiEngine repo].

import torch
import MinkowskiEngine as ME

data_batch_0 = [
    [0, 0, 2.1, 0, 0],  #
    [0, 1, 1.4, 3, 0],  #
    [0, 0, 4.0, 0, 0]
]


def to_sparse_coo(data):
    # An intuitive way to extract coordinates and features
    coords, feats = [], []
    for i, row in enumerate(data):
        for j, val in enumerate(row):
            if val != 0:
                coords.append([i, j])
                feats.append([val])
    return torch.IntTensor(coords), torch.FloatTensor(feats)

def sparse_tensor_initialization():
    coords0, feats0 = to_sparse_coo(data_batch_0)
    coords0 = coords0.cuda()
    feats0 = feats0.cuda()
    coords0, feats0 = ME.utils.sparse_collate(coords=[coords0], feats=[feats0], device=coords0.device)
    A = ME.SparseTensor(coordinates=coords0, features=feats0)
    return A


if __name__ == '__main__':
    sparse_tensor_initialization()

I am using cuda version11.2

[chrischoy]

I cant replicate the error. Could you report the environment as specified on the issue template?

[victoryc]

Unfortunately, after I posted the issue, for other reasons, I had to upgrade pytorch to the the latest version (i.e., 1.8.1) which as I understand isn't compatible with MinkowskiEngine. So, I am not currently able to provide the complete and exact environment info corresponding to the state when I had the issue. Below is partial info about the environment.

• OS: Ubuntu 20.04 in WSL2 (Windows Subsystem for Linux 2) running on top of Windows 10
• Python version: 3.8.5
• Output of the following command. (If you installed the latest MinkowskiEngine, simply call MinkowskiEngine.print_diagnostics())
• ==========System==========
  Linux-5.4.91-microsoft-standard-WSL2-x86_64-with-glibc2.10
  DISTRIB_ID=Ubuntu
  DISTRIB_RELEASE=20.04
  DISTRIB_CODENAME=focal
  DISTRIB_DESCRIPTION="Ubuntu 20.04 LTS"
  3.8.5 (default, Sep 4 2020, 07:30:14)
  [GCC 7.3.0]
  ==========Pytorch==========
  1.7.1
  torch.cuda.is_available(): True
  ==========NVIDIA-SMI==========

[chrischoy]

I see. It seems like this was a pytorch 1.8 related issue. 1.8 compatibility will be fixed in the next version.

[chrischoy]

This is the pytorch 1.8.1 + CUDA 11 error on #330. Please refer to the #330 for updates.