Pytorch version upgrade >=2.2

[gmehtaa]

Pytorch version upgrade >=2.2
https://pytorch.org/docs/stable/index.html

[dsajdak]

This is on hold until the next CCR software release comes out

[purusoni]

Is there any timeline for the next release?

[dsajdak]

@puru-soni-04 The new release is in beta testing now. See here. We are working to build updated software packages for ccrsoft/2024.04 throughout the summer.

[purusoni]

So is there any timeline for the PyTorch module in the new release? if not, could you please share the easybuild recipe to install PyTorch module on ccrsoft/2024.04?

[dsajdak]

I'm sorry but I can't say for sure. There are many competing tasks on the summer to-do list. We don't have a new easybuild recipe for that yet as we haven't tried to build it. You can take the one we have for the previous version and attempt to modify it for a newer version either with the toolchains in 2023.01 or use the Easybuild example for the new software release 2024.04, currently in beta testing, and change the pytorch version. We have some documentation here on using Easybuild. However, building PyTorch can be challenging so either option may not work for you.

[dsajdak]

Just an update on this request as I know there are many following it:
ccrsoft/2024.04 release is available but only the compilers and toolchains are built right now. We’re working on building out the software available in this release. What we’ve been recommending to users who want more recent versions of pytorch and other AI/ML packages is to use the NVIDIA containers. We have some info on this in our documentation:
https://docs.ccr.buffalo.edu/en/latest/howto/containerization/
https://docs.ccr.buffalo.edu/en/latest/howto/containerization/#gpu-enabled-containers

This is probably your best bet going forward because we will only be offering one version of each software application in each software release and this is often not enough for the AI/ML workloads. These applications need to be tied to a specific CUDA version and NVIDIA drivers that match our systems so often the version we’re able to install is not the newest.

Along with additional documentation, some training workshops, and new container options, we also plan to provide a container repository to make pulling containers easier and faster on our systems. I know this doesn’t help you all today, but this is the future at CCR for these more complicated software packages so we recommend you consider containers for your workflow. Please feel free to submit tickets to ccr-help if you're trying to use the NVIDIA containers and are running into issues. We'd be happy to provide guidance.