About the version of mmcv corresponding to torch 2.0.0 and cuda118

1

Hello everyone!
I am trying from .iou3d import iou3d_nms_cuda and an error is raised as follows

iou3d_nms_cuda.cpython-38-x86_64-linux-gnu.so: undefined symbol: 
_ZNK3c1010TensorImpl36is_contiguous_nondefault_policy_implENS_12MemoryFormatE

From the site I know that there is no mmcv-full package in my environment. However, when I turn to Docs of mmcv-full, there is no corresponding version to torch2.0.0+cuda118, the environment in my computer.

Then I tried to pip install mmcv-full==1.7.1 which is the newest version, but that doesn’t work.

So what can I do to solve this problem? Or should I create a new environment with lower version of torch and cuda?

2

is_contiguous_nondefault_policy was removed in this PR in May 2022 and thus is not available anymore since torch>=1.12.0. I don’t know if the authors of the mmcv repository are rebuilding their binaries using their latest release, which might be compatible to the current PyTorch version, but based on their README it seems their latest pip wheel supports torch==1.8.0+cu102, so you might need to build mmcv from source.

1 Like
3

Now I’m going to create a virtual env to satisfy the version requirements.