CUDA Error for Blackwell (RTX5090) GPU

[lian-jiarong]

Hello, I encountered the error "RuntimeError: CUDA error: no kernel image is available for execution on the device". Are there any debug suggestions for Blackwell GPUs? My CUDA version is 12.8, and the torch version is 2.7.1. Thanks!

[QQAndYY-cyber]

I encountered the same problem as you; may I ask if you have found a solution?

[Jianhao-zheng]

Hi, I don't have a 5090 on hand to test. But I think you need to add:

'-gencode=arch=compute_120,code=sm_120',
'-gencode=arch=compute_120,code=compute_120',

in setup.py:

WildGS-SLAM/setup.py

Line 26 in 498c768

'-gencode=arch=compute_86,code=sm_86',

Please note that the same thing should be added to other setup.py of the thirdparty package.

In addition, since the CUDA version is very new, you probably also need to change something in the lietorch source code.

[panyangis]

python -c "import torch; import lietorch; import simple_knn; import diff_gaussian_rasterization; print(torch.cuda.is_available())" Traceback (most recent call last): File "<string>", line 1, in <module> File "/media/useradmin/B32C-9D8F/SLAM project/WildGS-SLAM/thirdparty/lietorch/lietorch/__init__.py", line 2, in <module> from .groups import LieGroupParameter, SO3, RxSO3, SE3, Sim3, cat, stack File "/media/useradmin/B32C-9D8F/SLAM project/WildGS-SLAM/thirdparty/lietorch/lietorch/groups.py", line 5, in <module> from .group_ops import Exp, Log, Inv, Mul, Adj, AdjT, Jinv, Act3, Act4, ToMatrix, ToVec, FromVec File "/media/useradmin/B32C-9D8F/SLAM project/WildGS-SLAM/thirdparty/lietorch/lietorch/group_ops.py", line 1, in <module> import lietorch_backends ImportError: libcudart.so.11.0: cannot open shared object file: No such file or directory
Hello, I have configured a 5060ti graphics card and used CUDA12.8. Now I have added the following code to WildGS-SLAM/setup.py and WildGS-SLAM/thirdparty

'-gencode=arch=compute_120,code=sm_120', '-gencode=arch=compute_120,code=compute_120',

But the error still occurs

python -c "import torch; import lietorch; import simple_knn; import diff_gaussian_rasterization; print(torch.cuda.is_available())" Traceback (most recent call last): File "<string>", line 1, in <module> File "/media/useradmin/B32C-9D8F/SLAM project/WildGS-SLAM/thirdparty/lietorch/lietorch/__init__.py", line 2, in <module> from .groups import LieGroupParameter, SO3, RxSO3, SE3, Sim3, cat, stack File "/media/useradmin/B32C-9D8F/SLAM project/WildGS-SLAM/thirdparty/lietorch/lietorch/groups.py", line 5, in <module> from .group_ops import Exp, Log, Inv, Mul, Adj, AdjT, Jinv, Act3, Act4, ToMatrix, ToVec, FromVec File "/media/useradmin/B32C-9D8F/SLAM project/WildGS-SLAM/thirdparty/lietorch/lietorch/group_ops.py", line 1, in <module> import lietorch_backends ImportError: libcudart.so.11.0: cannot open shared object file: No such file or directory
I sincerely ask you what source code needs to be modified

[Bo-Wing]

If you’ve found a solution, could you post the steps or key changes here? Thanks!

[towardsDLCV]

Hello, I’m also using an RTX 5090. The lietorch included in this repository is built for CUDA 11, so rebuilding it from the official lietorch GitHub repository (https://github.com/princeton-vl/lietorch).