Using ProteinMPNN on AMD GPUs at Pawsey

Installation

1. Clone the ProteinMPNN repository into your software directory:

Dependencies

ProteinMPNN requires PyTorch, which is available as an optimized system-wide module on Pawsey. To use it:

1. Check available PyTorch versions:

2. Load the appropriate PyTorch module:

Note: No additional conda environment setup is required as all dependencies are handled by the system module.

Running ProteinMPNN on Pawsey GPUs

SLURM Job Script Template

Create a job script (e.g., run_proteinmpnn.sh) with the following configuration. In particular, note the loading of the pytorch module and adding the correct srun parameters before each python task.

Important Parameters

• --partition=gpu: Specifies the GPU partition

• --nodes=1: Number of nodes to use

• --gres=gpu:1: Requests one GPU

• --account=${PAWSEY_PROJECT}-gpu: Your project's GPU account

• --time=1:00:00: Job time limit (adjust as needed)

Running Tasks

Each Python task in the script uses srun with specific GPU parameters:

• -N 1: One node

• -n 1: One task

• -c 8: 8 CPU cores per task

• --gres=gpu:1: One GPU per task

• --gpus-per-task=1: One GPU per task

• --gpu-bind=closest: Optimal GPU-CPU binding

Submit the Job

Submit your job to the SLURM scheduler:

Check job status:

Further Reading

For more details on running GPU workflows on Setonix, refer to Setonix GPU Partition Quick Start