Plugin start-up fails under Ubuntu 24.04 with CUDA 13

[Skelp]

issue description

• have CUDA 13.x installed
• install according to instructions
• start LichtFeld-Studio
• discover that the plugin failed to load

undefined symbol: cudaGetDriverEntryPointByVersion, version libcudart.so.12

temporary solution

resolved by changing all references of cu128 (torch related) to cu130 in the pyproject.toml file, remove uv.lock (unsure if it is required).

proposed long-term solution

detect the environment at hand to determine which CUDA variant to use during installation and / or allow the user to choose during installation

---

neofetch

OS: Ubuntu 24.04.4 LTS x86_64  
Kernel: 6.17.0-14-generic

nvidia-smi

NVIDIA-SMI 580.95.05              Driver Version: 580.95.05      CUDA Version: 13.0

nvcc --version

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2025 NVIDIA Corporation
Built on Wed_Aug_20_01:58:59_PM_PDT_2025
Cuda compilation tools, release 13.0, V13.0.88
Build cuda_13.0.r13.0/compiler.36424714_0

[shadygm]

i'm on cuda 13.1, but I don't see these issues.

Its usually backwards compatible, so I don't see why the issue should exist in the first place. Can you let me know if fully uninstalling the plugin from the ~/.lichtfeld/plugin directory and then attempting to install it again resolves the issue? I just pushed a change on main

[Skelp]

Seems like there was contradicting installations on my end. CUDA installations are hell to manage, and since I had both a manually installed CUDA 13.0 and an apt-installed nvidia-cuda-toolkit (CUDA 12.x), nvidia-smi reported CUDA 13 but nvcc reported CUDA 12.x.

[Skelp]

Okay, reopening this issue since this was not the cause of the problem it seems.
Paths are all pointing to CUDA 13. VCPKG downloads have been purged, LFS build-dir deleted and rebuild, plugin reinstalled via UI for sanity checking. Now I get the same issue regarding cudaGetDriverEntryPointByVersion. Once again, enforcing cu130 for pytorch solves the issue.
I have not a single clue as to why this happens - before, it was clear that NVCC reported a different version than NVIDIA-SMI. Now that they are synced up and paths are clean I don't see why pytorch-cu128 would fail, but it does regardless.

[shadygm]

@Skelp, the pytorch wheels should be self contained. Could you try deleting the whole ~/.lichtfeld directory, there are some venvs there that might be messing up the paths, tho it should not really happen. or perhaps your library paths are a bit mangled, check you LD_LIBRARY_PATH as well.

[Skelp]

Been there, done that. It's not the venv's directory fault.
LD_LIBRARY_PATH is correct as well, as I've considered all paths while cleaning.

[Skelp]

The wheel might be self-contained, but unless it ships with its own copy of libcudart, I kind of understand why it crashes. If it tries to access something only available for CUDA 12.8 (or at least not anymore above 12.8), then it seems logical it would die trying to call removed functions.

[shadygm]

I find that still quite odd, like I said, I'm on cuda 13.1, with nothing cuda 12.8 installed on my system yet I'm still able to run it. I'm fairly certain that the wheel is fully self-contained to avoid such issues. Did you find another solution than to change to cuda 13.0? Determining which cuda the system picks like you proposed is not really feasible from my side since LFS installs all of the dependencies based on the project toml file without any consideration for which cuda is installed.