PyTorch¶

PyTorch from the software stack:¶

PyTorch with gpu-support is available in the software stack. You need to load the following modules to enable PyTorch:

module load 2026 gpu openmpi py-torch

You can check for the availability of a gpu card with PyTorch with the following python script:

howmanygpus.py

import torch
cuda_avail = torch.cuda.is_available()
if cuda_avail:
    print("Torch CUDA is available")
    num_of_devices = torch.cuda.device_count()
    if num_of_devices:
        print("Number of CUDA devices: {}".format(num_of_devices))
        current_device = torch.cuda.current_device()
        current_device_id = torch.cuda.device(current_device)
        current_device_name = torch.cuda.get_device_name(current_device)
        print("Current device id: {}".format(current_device_id))
        print("Current device name: {}".format(current_device_name))
    else:
        print("No CUDA devices!")
else:
    print("Torch CUDA is not available!")

This script can be run with sbatch as follows:

howmanygpus.slurm

#!/bin/bash
#SBATCH --job-name="pytorch/howmanygpus"
#SBATCH --output=howmanygpus.out
#SBATCH --time=00:10:00
#SBATCH --ntasks=1
#SBATCH --cpus-per-task=1
#SBATCH --gpus-per-task=1
#SBATCH --partition=gpu-a100-small
#SBATCH --mem-per-cpu=1G
# make sure to add your account!
##SBATCH --account=<what>-<faculty>-<group>

module load 2026 gpu
module load openmpi
module load py-torch

srun python howmanygpus.py

If you have a more illustrative example that you would like to share, please post on mattermost or send it to info-DHPC@tudelft.nl.

Install your own PyTorch with a virtual environment¶

If you need a different PyTorch version than the one provided by the py-torch module above, you can build your own directly on a login node using a plain Python virtual environment together with the gpu stack:

module load 2026 gpu python
python -m venv env
source env/bin/activate
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130

Loading the gpu stack before creating the environment is what makes this work from a login node without a GPU card present, and the resulting env will run unmodified on the GPU compute nodes. See Installing GPU-enabled packages for more details on this workflow.

About the CUDA version

This pip-installed PyTorch bundles its own CUDA 13.0 runtime as pip dependencies — it does not need, and should not be combined with, a module load cuda. If you separately need to compile a custom CUDA extension with nvcc, note that the module system's cuda modules currently only go up to cuda/12.9, slightly behind this wheel's CUDA 13.0 runtime. In that case, use the cu126 index instead (https://download.pytorch.org/whl/cu126), which stays closer to what the module system's cuda toolkit versions can offer.

Install your own PyTorch with conda¶

Alternatively, you can install your own GPU-enabled PyTorch version with conda and pip in your /home directory, from a login node. To do so:

Install the miniconda3 package manager as described here,

Install PyTorch with GPU-support in an environment:

conda create -n mytorchgpuenv python=3.10.13
conda activate mytorchgpuenv
(mytorchgpuenv)$ pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu130

After this, all you need in your submission script is to activate the conda environment. No need to load extra modules. The same CUDA-version note above applies here too.