CUDA error on RTX 5090 (sm_120) – PyTorch in Docker image does not support new architecture

[BYALPERENK]

Hi,
I am running FastSurfer via the official Docker image (deepmi/fastsurfer:latest) on a system with an NVIDIA GeForce RTX 5090 (compute capability sm_120).

Segmentation fails with the following error when using CUDA:

UserWarning: NVIDIA GeForce RTX 5090 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.

CUDA error: no kernel image is available for execution on the device
ERROR: FastSurfer asegdkt segmentation failed.

This suggests that the PyTorch version bundled in the current Docker image was built without support for sm_120, so GPU execution is not possible on RTX 5090.
Running with --device cpu works as expected.

Is there a plan to update the Docker image with a newer PyTorch/CUDA build that supports RTX 5090 (Blackwell, sm_120)?

Thanks.

[dkuegler]

Hi,
we are aware that the current docker image is not supported by 50-series GPUs. We are working on a new release of FastSurfer, which will support 50-series GPUs. This essentially is because nVIDIA dropped support fro older CUDA version with this new generation, but the pytorch version we tested everything with does not support those newer CUDA versions.

So unfortunately, there will likely be no official FastSurfer 2.4.* image that supports 50-series GPUs. And fortunately, there will be a new FastSurfer version altogether soon :-).

Cheers, David