I ssues install scNodes on RedHat Linux

[cisbiosbg]

Dear Developers

I am slightly confused with the installation instructions & having issues for scNodes:
OS: Redhat 9; GCC 11.4.1; Python 3.10.15; Cuda 11.8
Note: Tried with cuda 12.4 but seems to fail in step 2 below - as it seems to depend on Cuda 11. Error was:
'...W tensorflow/stream_executor/platform/default/dso_loader.cc:64]
Could not load dynamic library 'libcudart.so.11.0'; dlerror: libcudart.so.11.0...'

I am trying to install on a head-node of a HPC farm, for general use.

This Part works (I can import scNodes e.g., python -c 'import scNodes')
1: pip install scNodes
However, with
2: python -m scNodes install
I get errors:
/somewhere/scNodes/lib/python3.10/site-packages/glfw/init.py:917: GLFWError: (65543) b'GLX: Forward compatibility requested but GLX_ARB_create_context_profile is unavailable'
warnings.warn(message, GLFWError)
Traceback (most recent call last):
...
Exception: Could not initialize GLFW window.

I also get errors trying to install gpufit directly, with error 'error: invalid command 'bdist_wheel'. Tried different versions of wheel (0.30.0-0.45.1)

Any suggestions welcome.

Regards
Suhail

[mgflast]

Hi Suhail,

Thanks for sharing the issue! These seem like 3 separate problems... sorry about that.

1. CUDA: I've just pushed an edit to fix the cuda error. Could you try to install that from here? pip install git+https://github.com/bionanopatterning/scNodes.git
2. GPUFIT: the authors of gpufit only distributed binaries for windows, which the python -m scNodes install command tries to install. For unix you would need to compile it from source, see instructions here: https://gpufit.readthedocs.io/en/latest/installation.html
3. GLFW window: this is somewhat common on HPC nodes that do not have full display support. Were you using x11 forwarding when you connected to the node? (ssh -X user@node). If yes and the problem still occurs, you could try to force the system to use software rendering instead:

export MESA_GL_VERSION_OVERRIDE=3.3
export LIBGL_ALWAYS_SOFTWARE=1
python -m scNodes install

Let me know if that helps. Best,
Mart

[mgflast]

one more thing, even on HPC nodes that do technically have display support, the OpenGL drivers are often very old and don't support the required glfw functions. In that case software rendering is also a possible workaround, though it may introduce quite a high latency.

[cisbiosbg]

Mart Hi
Thank you very much for your prompt reply. I will look at your suggestions 1st thing on Monday and keep you updated. Will also need to look at how this this software can be used on a HPC farm with GPU comp nodes. The Head node is simply a CPU box.
Regards
Suhail

[mgflast]

Hi Suhail,

I also had this glfw issue on our HPC nodes when I moved to a new institute and I have to say I never really figured out how to fix it... so instead I've worked around it. There are some workarounds possible but whether they are practical depends on what features of the software you would like to use.
If it is mainly the fluorescence data processing and using gpufit, I would think that a single GPU is sufficient, so in that case a somewhat powerful desktop PC could work. Or if people use Warp for cryoET, there might also be a dedicated Warp Windows PC around (old Warp used to be windows only, so some labs got dedicated machines for it...).
If it is the segmentation part of the software, I have recently 'decoupled' the GUI from the GPU-heavy processing in the standalone version of it (github.com/bionanopatterning/Ais). So I run the GUI for the annotation on my Windows desktop PC, then when I want to train a network or use it for inference I use the command line interface options on the HPC cluster.

Maybe an approach like that could work - I would be happy to think along.

Have a nice weekend!