Machine learning framework for Python.
5.10.2022 Due to Puhti's update to Red Hat Enterprise Linux 8 (RHEL8), the number of fully supported PyTorch versions has been reduced. Previously deprecated conda-based versions have been removed. Please contact our servicedesk if you really need access to older versions.
5.5.2022 Due to Mahti's update to Red Hat Enterprise Linux 8 (RHEL8), the number of fully supported PyTorch versions has been reduced. Please contact our servicedesk if you really need access to other versions.
4.2.2022 All old PyTorch versions which were based on direct Conda installations have been deprecated, and we encourage users to move to newer versions. Read more on our separate Conda deprecation page.
Currently supported PyTorch versions:
All modules include PyTorch and related libraries with GPU support via CUDA.
Versions marked with "(x)" are based on old Red Hat Enterprise Linux 7
(RHEL7) images, and are no longer fully supported. In particular MPI
and Horovod are not expected to work anymore with these modules. If
you still wish to access these versions, you need to enable old RHEL7
module use /appl/soft/ai/rhel7/modulefiles/.
If you find that some package is missing, you can often install it yourself with
pip install --user. See our Python
more information on how to install packages yourself. If you think that some
important PyTorch-related package should be included in the module provided by
CSC, please contact our servicedesk.
All modules are based on containers using Apptainer (previously known
as Singularity). Wrapper scripts have been provided so that common
commands such as
pip3 should work as
normal. For other commands, you need to prefix them with
apptainer_wrapper exec, for example
huggingface-cli. For more information, see CSC's general
instructions on how to run Apptainer
If you are new to using machine learning on CSC's supercomputers, please read our new tutorial Getting started with machine learning at CSC, which covers how to run a simple PyTorch project on Puhti using the web user interface.
PyTorch is BSD-style licensed, as found in the LICENSE file.
To use this software on Puhti or Mahti, initialize it with:
module load pytorch
to access the default version, or if you wish to have a specific version (see above for available versions):
module load pytorch/1.12
Please note that the module already includes CUDA and cuDNN libraries, so there is no need to load cuda and cudnn modules separately!
This command will also show all available versions:
module avail pytorch
To check the exact packages and versions included in the loaded module you can run:
Note that login nodes are not intended for heavy computing, please use slurm batch jobs instead. See our instructions on how to use the batch job system.
Example batch script
Example batch script for reserving one GPU and 1/4 of the available CPU cores in a single node:
#!/bin/bash #SBATCH --account=<project> #SBATCH --partition=gpu #SBATCH --ntasks=1 #SBATCH --cpus-per-task=10 #SBATCH --mem=64G #SBATCH --time=1:00:00 #SBATCH --gres=gpu:v100:1 module load pytorch/1.12 srun python3 myprog.py <options>
#!/bin/bash #SBATCH --account=<project> #SBATCH --partition=gpusmall #SBATCH --ntasks=1 #SBATCH --cpus-per-task=32 #SBATCH --time=1:00:00 #SBATCH --gres=gpu:a100:1 module load pytorch/1.11 srun python3 myprog.py <options>
Please read the section on Efficient GPU utilization in our Machine learning guide to learn how to use the GPU efficiently.
Big datasets, multi-GPU and multi-node jobs
If you are working with big datasets, or datasets that contain a lot of files, please read the data section of our Machine learning guide. In particular, please do not read a huge number of files from the shared file system, use fast local disk or package your data into larger files instead!
For multi-GPU and multi-node jobs we recommend using the PyTorch Distributed Data-Parallel framework. You can read more about this and find examples of how to use PyTorch DDP on CSC's supercomputers in the Multi-GPU and multi-node section of our Machine learning guide