Tried to allocate 131060.03 GiB

[KennyFu1997]

Hello,

Thank you for the work!
I am using GPS_Gaussian which uses your library for Gaussian rendering. The compilation with pip install -e submodules/diff-gaussian-rasterization is OK. But the problem starts from here num_rendered, color, radii, geomBuffer, binningBuffer, imgBuffer, invdepths = _C.rasterize_gaussians(*args), it reports to allocate 131TB memory which is weird.

Platform: Linux Ubuntu 24.04, the cuda driver version is 12.7 reported by typing "nvidia-smi"
The cuda runtime version on my machine is 12.8 /usr/local/cuda-12.8/bin/nvcc and I install torch with cuda 12.8. I believe that this must be version mismatching problem. Previously I installed torch with cuda 12.4, and the problem was instead "segmentation fault". Now the segfault is gone but this new problem comes.
Any advice would be appreciated!

[KennyFu1997]

Solved. use cuda toolkit version of 11.8 in conda env instead.