Internal compiler error: Segmentation fault

[mertkiray]

Hello,

While building with python project.py build

I got the following error:

/usr/include/c++/10/chrono:473:154: required from here /usr/include/c++/10/chrono:428:27: internal compiler error: Segmentation fault 428 | _S_gcd(intmax_t __m, intmax_t __n) noexcept | ^~~~~~

My OS: Ubuntu 21.04
Shipped gcc version:10.3
Installed gcc version: 9.3.0
nvcc version: 11.2
Nvidia-smi:

Thanks for the help

[chengzhag]

Hi.

It looks like that the problem happens when building the LDIF kernel provided by https://github.com/google/ldif. Thus we cannot provide a direct solution.
However, the LDIF kernel is not a must-have. You can still run our code following instructions in issue #1 without it.

[mertkiray]

One quick idea:
New RTX 30 series nvidia gpus require cuda toolkit >= 11.1 and from pytorch site
NOTE: 'nvidia' channel is required for cudatoolkit 11.1 conda install pytorch torchvision torchaudio cudatoolkit=11.1 -c pytorch -c nvidia
But in the environment.yml this requirement is not met so the gpu cannot be utilized.
Can we do a quick check?

[chengzhag]

The compiling of LDIF kernel might be using the nvcc installed on the system. But checking the cudatoolkit version of the conda environment might be a good starting point if any problem happens when running our code with RTX GPU.

If the problem of compiling remains, you can always try the solution mentioned in issue #1. According to the observation of some other fellows, the CUDA kernel provides no improvement on RTX GPU thus might be not requisite.

[Fizmath]

Try on with UBUNTU 20.04.
NVIDIA has tested its latest GPU's up to CUDA 11.3.0 with UBUNTU 20.04 and UBUNTU 18.04.