Installation issue

[xtalfire]

Hi,
I tried to pull the docker/apptainer as indicated from the tutorial, but it failed, the link is not responsive.
Then I tried to git clone and create a conda environment but it seems there is a CUDA lib /system mismatch.
" libcusparse.so.12: undefined symbol: __nvJitLinkComplete_12_4, version libnvJitLink.so.12 "

I tried also to build an apptainer from the Singularity file inside the git folder, but it also fails,
" RuntimeError: CUDA error: out of memory
CUDA kernel errors might be asynchronously reported at some other API call, so the stacktrace below might be incorrect.
For debugging consider passing CUDA_LAUNCH_BLOCKING=1
Compile with TORCH_USE_CUDA_DSA to enable device-side assertions. "

I wonder is it possible to update the link so I can try to get the apptainer directly?
Thanks.

[MihinP]

Hello,

Thanks for trying out PartiNet and raising this issue.

A few quick questions to help us debug:

1. Are you on a shared HPC cluster or a personal workstation/desktop?
2. If on a cluster, are you running commands from a login node or a compute node?

Container link:
We've verified this is working on our end (fresh 3.1 GiB pull succeeded), so this is likely a network/firewall issue.
Try running curl -sI https://ghcr.io and report back the output. A correct working output should look something like the following:

HTTP/2 405
content-type: application/json
docker-distribution-api-version: registry/2.0
strict-transport-security: max-age=63072000; includeSubDomains; preload
date: Mon, 27 Apr 2026 00:31:21 GMT
content-length: 78
x-github-request-id: AC31:1B0133:4EAD48:9D578B:69EEAE59

If not, your cluster may block access to the container registry and you will need to contact a system admin or research computing support to look at this.

---

Conda CUDA mismatch

This error means the PyTorch build in your conda environment was compiled against a different CUDA version than the one available on your system. To help fix this can you please connect to/ssh into a GPU node (if you are on a cluster) and run nvidia-smi. An example output is the following:

Mon Apr 27 10:10:07 2026
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 570.86.15              Driver Version: 570.86.15      CUDA Version: 12.8     |
|-----------------------------------------+------------------------+----------------------+
| GPU  Name                 Persistence-M | Bus-Id          Disp.A | Volatile Uncorr. ECC |
| Fan  Temp   Perf          Pwr:Usage/Cap |           Memory-Usage | GPU-Util  Compute M. |
|                                         |                        |               MIG M. |
|=========================================+========================+======================|
|   0  NVIDIA A30                     On  |   00000000:17:00.0 Off |                    0 |
| N/A   22C    P0             26W /  165W |       1MiB /  24576MiB |      0%      Default |
|                                         |                        |             Disabled |
+-----------------------------------------+------------------------+----------------------+

+-----------------------------------------------------------------------------------------+
| Processes:                                                                              |
|  GPU   GI   CI              PID   Type   Process name                        GPU Memory |
|        ID   ID                                                               Usage      |
|=========================================================================================|
|  No running processes found                                                             |
+-----------------------------------------------------------------------------------------+

The top right corner contains the CUDA version. Once we have this we can give you a specific pip install command to get around this error. In the meantime we will look into the install dependencies to prevent this from happening to other users

If you are on a cluster you may also have to load a CUDA module separately. Try running module avail CUDA and report the output.

---

Apptainer build OOM

The CUDA error: out of memory during the Apptainer build is most likely caused by a GPU call being triggered at install/import time during the build. The Singularity definition file does not require a GPU to build. Try building on a CPU-only node, or force CPU-only mode by running:

CUDA_VISIBLE_DEVICES="" apptainer build partinet.sif Singularity

---

See how you go and let us know the outputs of these commands.

Cheers,
Mihin