Closed mrclovvn closed 3 years ago
JohnRomanelis
commented
3 years ago Hey @mrclovvn! I had the same error. The problem is with the cuda version you have. I suppose you have installed CUDA 11.0 as it is the latest version you see as an option when installing pytorch. I tried installing CUDA 11.1, then installing pytorch with cuda 11.0 and the problem is gone! I got some warnings during installation but i was able to install similar libraries that threw the same error.
mrclovvn
commented
3 years ago Hey @mrclovvn! I had the same error. The problem is with the cuda version you have. I suppose you have installed CUDA 11.0 as it is the latest version you see as an option when installing pytorch. I tried installing CUDA 11.1, then installing pytorch with cuda 11.0 and the problem is gone! I got some warnings during installation but i was able to install similar libraries that threw the same error.
Thanks @JohnRomanelis! A few days ago when I saw README.md has updated the requirements about the supported CUDA version then I tried CUDA11.1, and there is indeed no any problem. The CUDA11.1 was not the supported version in the previous requirements, so I have not tried it before.
Hi @chrischoy
I got a new 3090 recently but when I try to install the latest version of ME I encountered some problem. It reported an error that:
nvcc fatal: Unsupported gpu architecture 'compute_86',so I’m sorry to bother you here to confirm if it’s not supported 3090 yet?
And I have tried to set the variable as ‘export TORCH_CUDA_ARCH_LIST="8.0" ’, then the ME can be installed successfully but I get a new error when I debug the FCGF benchmark_3dmatch.py:
device-side assert triggered at /home/.../MinkowskiEngine-0.5/src/kernel_region.hpp:330in MinkowskiConvolution.py line 72. However, the weird thing is if I just run benchmark_3dmatch.py eveything is OK.If you have any suggestions I would appreciate it. Thank you.