PyTorch CUDA incompatibility with NVIDIA RTX 5070 Ti

Support
1

Hello everyone,

I’m currently using a 3D Slicer (version 5.9.0-2025-06-03) on Ubuntu Linux. I have an NVIDIA GeForce RTX 5070 Ti GPU installed.

When launching Slicer, I’m consistently seeing the following UserWarning in the Python Interactor:

NVIDIA GeForce RTX 5070 Ti with CUDA capability sm_120 is not compatible with the current PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_50 sm_60 sm_70 sm_75 sm_80 sm_86 sm_90.
If you want to use the NVIDIA GeForce RTX 5070 Ti GPU with PyTorch, please check the instructions at https://pytorch.org/get-started/locally/

My understanding is that while Slicer detects my GPU, the PyTorch version bundled does not support the sm_120 (compute capability 12.0) architecture of my RTX 5070 Ti.

This means that Slicer modules are unable to access or utilize my GPU for accelerated processing, and consequently, will fall back to CPU execution, which is slower.

Is there anyone who can help resolve this issue so that the Slicer modules can run on this GPU?

Thank you so much.

2

As far as I can tell only 12.8 series support 5070Ti. So uninstall your pytorch (I assume you are using PyTorch utils?) restart slicer, then instead of automatic search for compatible versions, choose specifically cu12.8 and try again.

3

Running on a 5090 - had to update Pytorch to 12.8 from within slicer python console. Around a 4 GB download so takes a while. SM_120 refers to the 50 series GPUs I believe, so 12.8 was the only supported version.

4

@Mark_Ryan You were successful with this on Windows? Or were you using a Linux distro? Rodrigo appears to be using Ubuntu 24.04 per:

5

This was from within windows, but everything downloaded within python shell in Slicer. But I needed ChatGPT to get me that far, so can’t speak for the Linux thing.

6

Ok thanks for the update. Yes, Rodrigo’s issues exists because of using the Slicer linux package which is currently built based on CentOS 7 and glibc 2.17 and there is not a Pytorch 2.7.x whl that is compatible with it.

Pytorch 2.6.0 for example has manylinux1 (CentOS 5 based, glibc 2.5) based whls which can be run by the Slicer linux package, however Pytorch 2.7.0 has manylinux_2_28 (glibc 2.28) based whls which is too new of a glibc to match what Slicer provides for the linux package.