From d529eee983fec665e1690b09f4be1b0388379169 Mon Sep 17 00:00:00 2001
From: Anja Virkkunen <anja.virkkunen@aalto.fi>
Date: Wed, 3 Jun 2026 12:51:02 +0300
Subject: [PATCH 1/2] PyTorch container example for ARM GPUs

---
 triton/usage/gracehopper.rst | 92 ++++++++++++++++++++++++++++++++++++
 1 file changed, 92 insertions(+)

diff --git a/triton/usage/gracehopper.rst b/triton/usage/gracehopper.rst
index 04c94eb6e..8bb0ade58 100644
--- a/triton/usage/gracehopper.rst
+++ b/triton/usage/gracehopper.rst
@@ -76,3 +76,95 @@ The hello-world.cu could be e.g.::
     return 0;
   }
 
+Building a simple pytorch environment
+-------------------------------------
+
+Nvidia provides ARM containers for PyTorch, which you can use as a starting point for your own containers.
+This example shows how you can extend such a container by installing additional packages from pip.
+
+A new PyTorch container is built each month, you can browse the selection
+in the `Nvidia PyTorch catalog <https://catalog.ngc.nvidia.com/orgs/nvidia/containers/pytorch>`__.
+The following container definition file selects the February 2026 Nvidia PyTorch container with PyTorch
+version 2.11 as a starting point:
+
+.. code-block:: none
+
+  Bootstrap: docker
+  From: nvcr.io/nvidia/pytorch:26.02-py3
+
+  %post
+    pip install transformers==4.57.6 pyyaml==6.0.1
+
+  %help
+    An apptainer image based on Nvidia's PyTorch container with ARM CPU architecture.
+
+    The bootstrapped container runs on Ubuntu 24.04, and contains CUDA version 13.1,
+    OpenMPI 4.1.7, Python 3.12 and PyTorch 2.11.
+
+    This image extends the bootstrapped PyTorch container with transformers package
+    from HuggingFace.
+
+    PyYAML is a package required by transformers that has already been installed by the
+    operating system package manager in the container. Pip will try to update PyYAML
+    (to 6.0.3 at the time this image was created), which will fail the build because
+    pip cannot change packages installed by the system package manager.
+    Thus, PyYAML has to be pinned to the version used by the system package manager.
+
+
+You can add other packages you need to the ``pip install`` command in the container definition above.
+We also recommend documenting your container in the ``%help`` section
+(which packages you have added and why).
+Save your container definition to a file (here we will use ``pytorch-transformers-arm.def``) and
+track it with version control to back it up for reproducibility.
+
+Next, you need to build the container image (SIF-file) from the definition file.
+Building needs to happen on an ARM-device, which you can achieve for example with
+the following sbatch script ``build-pytorch-transformers-arm-container.sh`` like so:
+
+.. code-block:: slurm
+
+  #!/bin/bash
+  #SBATCH --job-name=build-arm-container
+  #SBATCH --partition=gpu-grace-h200-141g
+  #SBATCH --cpus-per-task=4
+  #SBATCH --gpus=1
+  #SBATCH --time=01:00:00
+  #SBATCH --mem=128G
+
+  # You can replace $WRKDIR with $PWD to create the cache in your current working dir
+  mkdir -p "$WRKDIR/apptainer_cache"
+  export APPTAINER_CACHEDIR="$WRKDIR/apptainer_cache"
+
+  apptainer build pytorch-transformers-arm.sif pytorch-transformers-arm.def
+
+After you have successfully built your container, you can start using it in your scripts.
+Here is a simple example of how to run an imaginary Python training script with two arguments
+using the container:
+
+.. code-block:: sh
+
+  #!/bin/bash
+  #SBATCH --job-name=train-script
+  #SBATCH --partition=gpu-grace-h200-141g
+  #SBATCH --cpus-per-task=8
+  #SBATCH --gpus=1
+  #SBATCH --time=04:00:00
+  #SBATCH --mem=256G
+
+  # The --nv argument makes the GPU available within the container
+  apptainer exec --nv pytorch-transformers-arm.sif \
+    python train_script.py \
+      --arg1 foo \
+      --arg2 bar
+
+You simply need to prepend calls to your scripts with the apptainer exec command.
+For a more comprehensive tutorial on apptainer, please see
+`the third lesson <https://coderefinery.github.io/hpc-containers/>`__ of our
+`Tuesday Tools & Techniques for HPC (TTT4HPC) course <../../training/scip/ttt4hpc-2024.rst>`__.
+Just keep in mind when reading the lesson that it assumes x86 architecture instead of ARM,
+so adjust the examples in the tutorial to use ARM.
+In other words, be sure to select an ARM container as your starting point,
+and run the building script on ARM hardware as shown above.
+And if you want or need any help setting up your ARM containers,
+you can always join `SciComp garage <../../help/garage.rst>`__ for help.
+

From 89a32529a37579ae88962f71a5953a6ad186a764 Mon Sep 17 00:00:00 2001
From: Anja <Vinye@users.noreply.github.com>
Date: Fri, 26 Jun 2026 08:31:04 +0300
Subject: [PATCH 2/2] Fix typo

---
 triton/usage/gracehopper.rst | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/triton/usage/gracehopper.rst b/triton/usage/gracehopper.rst
index 8bb0ade58..65102d3f0 100644
--- a/triton/usage/gracehopper.rst
+++ b/triton/usage/gracehopper.rst
@@ -76,7 +76,7 @@ The hello-world.cu could be e.g.::
     return 0;
   }
 
-Building a simple pytorch environment
+Building a simple PyTorch environment
 -------------------------------------
 
 Nvidia provides ARM containers for PyTorch, which you can use as a starting point for your own containers.