Pytorch

Installed versions

Resource	Version
Dardel/cpe23.12	2.0.1

General information

PyTorch is an open-source machine learning library for Python, based on Torch. PyTorch provides two high-level features: - Tensor computation (like NumPy) with strong GPU acceleration - Deep neural networks built on a tape-based autodiff system

How to use

PyTorch is installed as a singularity container at PDC. The container includes PyTorch 2.0.1 with support for AMD GPUs using Rocm-5.7. In order to run the PyTorch container, first allocate a GPU node. Then, load the PDC and the singularity modules.

ml add PDC
ml add singularity/4.1.1-cpeGNU-23.12

PDC containers are placed at /pdc/software/sing_hub. They can also be reached by invoking PDC_SHUB.

ls $PDC_SHUB

Now that singularity is loaded, we can for example open a shell session within the container with: singularity shell --rocm -B /cfs/klemming /pdc/software/resources/sing_hub/rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1 The --rocm flag tells singularity to use the GPUs, and the -B flag tells singularity to mount /cfs/klemming into the container, thereby, making it possible to access files placed outside it. After the command, a prompt like this will appear:

Singularity>

Which means that we are inside the container, and now we can use python* to run our PyTorch scripts from the terminal. PyTorch models can also be run directly without the need of opening a shell in the container using the following command:

singularity exec --rocm -B /cfs/klemming /pdc/software/resources/sing_hub/rocm5.7_ubuntu22.04_py3.10_pytorch_2.0.1 python3 <pytorch_script.py>

Installing additional Python packages

There are some restrictions regarding singularity on Dardel. Users cannot write into singularity containers, and therefore, users cannot install additional python packages directly into the container. So if you need additional python packages, you can add them via one of these two following methods: - Download the PyTorch singularity container to your local machine where you have a local installation of singularity. Then, update the image there installing the packages needed, and afterwards upload the image to Dardel again. - Install the additional python packages directly on Dardel, but outside the singularity container. For example, in your home directory (the default behaviour with pip), or in your project directory using the --target flag with the pip command. Remember that if you install additional python packages outside the default python directories, you need to update the PYTHONPATH variable with the path where those new packages are before you run the container.

When installing new packages it is advisable to add the condition numpy==1.21.2 to pip install in order to prevent pip from upgrading the version of numpy in the container, which would cause compatibility problems. Example:

pip install matplotlib numpy==1.21.2