Pytorch¶

How to use on Arrhenius HPC¶

A limited number of PyTorch versions are available on Arrhenius through modules.

module avail PyTorch

if you need versions beyond what is available as module, the main recommendation is to use Apptainer containers instead, you can use e.g. either NVIDIAs NGC container catalog or the official images on dockerhub.

Internode communication with Slingshot¶

If you run PyTorch accross nodes you need to make sure that NCCL makes use of the Slingshot interconnect. Otherwise it will use the much slower ethernet connection.

With modules you do this by loading a NCCL plug-in with the module AWS-OFI-NCCL:

module avail AWS-OFI-NCCL

if you are using containers for this, note that you need to build this plug-in inside the container.

Feel free to contact support if you have any questions or concerns related to this.