Possible Typo in PyTorch Version (2.12 instead of 2.1.2?) & ImportError Issue
Hi, thank you for sharing this great project!
I noticed that in the README file, it is mentioned:
We've tested on RTX 4090 GPU with PyTorch version 2.12.
However, the latest official PyTorch version as of now is 2.5, and I couldn't find a version 2.12 in the release history. I was wondering if this was a typo and whether you actually meant 2.1.2 instead of 2.12?
To verify, I tried running the project with PyTorch 2.1.2, but encountered the following error:
ImportError: /data/khlee01/anaconda3/envs/Ex4DGS/lib/python3.9/site-packages/diff_gaussian_rasterization_df/_C.cpython-39-x86_64-linux-gnu.so: undefined symbol: _ZN3c1021throwNullDataPtrErrorEv
This seems to be related to a compatibility issue with diff_gaussian_rasterization_df and PyTorch.
Could you clarify which PyTorch version was actually used and whether there are any specific dependencies required to resolve this error?
Thank you in advance!
Hi, thanks for reminding me of this typo! As you mentioned, I've used PyTorch 2.1.2. To the best of my knowledge, most errors related to cpython come from the unmatched CUDA compiler versions. I recommend checking your CUDA compiler (use nvcc -v) and PyTorch's CUDA version. In my experience, older / newer PyTorch versions can run my code.