Hello,

 

It looks like you are correct, and the version of torch you are running is a bit old.

 

Can you check you torch version?

In the dev branch, we are currently working with `torch= 2.1.2+cu121`, which has support for the sm_90 compute capability.

>>> import torch

>>> torch.cuda.get_arch_list()

['sm_50', 'sm_60', 'sm_70', 'sm_75', 'sm_80', 'sm_86', 'sm_90']

 

You will almost certainly need to update your torch version, however I would strongly advise against updating to a newer point version of torch, as the utilities have not been tested against all the newer versions (i.e. the newest 2.4.0 has not been tested).

 

Best,

Jackson

 

From: freesurfer-bounces@nmr.mgh.harvard.edu <freesurfer-bounces@nmr.mgh.harvard.edu> on behalf of AmirHussein Abdolalizadeh <amirhussein.a@gmail.com>
Date: Wednesday, September 25, 2024 at 5:47 AM
To: Freesurfer support list <freesurfer@nmr.mgh.harvard.edu>
Subject: [Freesurfer] Repost: NextBrain Segmentation PyTorch-GPU Compatibility Error

        External Email - Use Caution        

Hi,

 

I am reposting an issue with NextBrain that I asked about a few months back: MailScanner has detected a possible fraud attempt from "secure-web.cisco.com" claiming to be https://www.mail-archive.com/freesurfer@nmr.mgh.harvard.edu/msg76705.html

 

"NVIDIA H100 with CUDA capability sm_90 is not compatible with the current

PyTorch installation.
The current PyTorch install supports CUDA capabilities sm_37 sm_50 sm_60
sm_70 sm_75 sm_80 sm_86."

 

There is a PyTorch incompatibility with the newer H100 nVidia GPUs. I had the same issue with FastBrain which was resolved by updating their torch dependencies in their release. See the discussion here:

 

I think this can be a solution for this issue as well since many HPCs are moving towards these kinds of GPUs.

 

Best,

Amir