From ac329336e8d86845f0ab81bde00e3a8a5e56caaa Mon Sep 17 00:00:00 2001
From: Mateo Langston Smith <mlangstonsmith@smu.edu>
Date: Wed, 14 Oct 2026 11:52:49 -0500
Subject: [PATCH 01/10] Add GPU for testing

---
 ood/rstudio/form.yml.erb           |  8 +++++++-
 ood/rstudio/template/script.sh.erb | 23 ++++++++++++++++++++++-
 2 files changed, 29 insertions(+), 2 deletions(-)

diff --git a/ood/rstudio/form.yml.erb b/ood/rstudio/form.yml.erb
index e8e1d62b..e30f66a6 100644
--- a/ood/rstudio/form.yml.erb
+++ b/ood/rstudio/form.yml.erb
@@ -5,6 +5,9 @@ require 'open3'
 require_relative "CheckQueues"
 require_relative "CheckAccounts"
 
+# testing: GPU-enabled custom container (bypasses the R module)
+gpu_tester = ["mlangstonsmith"].include?(User.new.name)
+
 -%>
 
 ---
@@ -65,6 +68,9 @@ attributes:
       <%- if CheckAccounts.Accounts.any? { |word| word.include?("kbnk") } | CheckAccounts.Accounts.any? { |word| word.include?("jlagrone") }-%>
       - [ "R 4.5.1 (custom kbnk)", "kbnk/R/4.5.1" ]
       <%- end -%>
+      <%- if gpu_tester -%>
+      - [ "GPU test container (rocker-ml-cuda)", "custom" ]
+      <%- end -%>
   environment:
     widget: text_area
     label: "Custom Environment Settings"
@@ -80,7 +86,7 @@ attributes:
       [SLURM partition]: https://southernmethodistuniversity.github.io/hpc_docs/about/queues.html#m3-queues
     options:
     <%- CheckQueues.Queues.each do |q| -%>
-      <%- if q[4].to_i == 0 -%> #disable gpu queues
+      <%- if q[4].to_i == 0 || gpu_tester -%> #disable gpu queues (except for testers)
       - ["<%= q[0] %>", "<%= q[0] %>",
         data-max-bc-num-hours: "<%= q[1] %>",
         data-max-bc-num-cores: "<%= q[2] %>",
diff --git a/ood/rstudio/template/script.sh.erb b/ood/rstudio/template/script.sh.erb
index 7606a8f7..4e1df900 100755
--- a/ood/rstudio/template/script.sh.erb
+++ b/ood/rstudio/template/script.sh.erb
@@ -12,6 +12,17 @@ env
 <%= context.environment.encode(universal_newline: true) %>
 <%- end -%>
 
+<%- if context.r_module == "custom" -%>
+# testing: GPU-enabled custom container, set what the R module would
+module load apptainer
+export CONTAINER_IMAGE=/lustre/scratch/client/users/mlangstonsmith/rocker-ml-cuda.sif
+export CONTAINER_RSERVER=/usr/lib/rstudio-server/bin/rserver
+export CONTAINER_RSESSION=/usr/lib/rstudio-server/bin/rsession
+export CONTAINER_R=/usr/bin/R
+export R_LIBS_USER=${HOME}/R/rocker-ml-cuda
+mkdir -p ${R_LIBS_USER}
+unset XDG_RUNTIME_DIR
+<%- else -%>
 # load the requested module
 FORM_MODULE=<%= context.r_module %>
 EXTENSION="${FORM_MODULE##*.}"
@@ -27,6 +38,7 @@ then
 else
   module load ${FORM_MODULE}
 fi
+<%- end -%>
 
 #
 # Start RStudio Server
@@ -53,6 +65,15 @@ sed 's/^ \{2\}//' > "${RSESSION_WRAPPER_FILE}" << EOL
 
   # make sure the container libR is visible inside container
   export LD_LIBRARY_PATH="/usr/local/lib/R/lib${LD_LIBRARY_PATH:+":$LD_LIBRARY_PATH"}"
+<%- if context.r_module == "custom" -%>
+
+  # environment.txt and the line above replace the container's paths, so
+  # restore libR (apt-installed in this image), the host driver libs bound
+  # by --nv, and the image's python venv (CUDA comes from pip wheels there)
+  export LD_LIBRARY_PATH="/usr/lib/R/lib:/.singularity.d/libs:\${LD_LIBRARY_PATH}"
+  export PATH="/opt/venv/bin:\${PATH}:/usr/lib/rstudio-server/bin/quarto/bin"
+  export CUDA_MODULE_LOADING="\${CUDA_MODULE_LOADING:-LAZY}"
+<%- end -%>
 
   # Log all output from this script
   export RSESSION_LOG_FILE="${PWD}/rsession.log"
@@ -195,7 +216,7 @@ unset APPTAINER_BIND
 
 export APPTAINERENV_XDG_CONFIG_DIRS="/etc/"
 export APPTAINERENV_RSTUDIO_CONFIG_DIRS=$RSTUDIO_CONFIG_DIRS
-apptainer exec \
+apptainer exec <%= context.bc_num_gpus.to_i > 0 ? "--nv" : "" %> \
   -B ${TMPDIR}:/tmp \
   -B ${TMPDIR}/run:/run \
   -B ${TMPDIR}/var/lib:/var/lib/rstudio-server \

From b9fba9e8756c004cc5b42cafd176b484f3de0cff Mon Sep 17 00:00:00 2001
From: Mateo Langston Smith <mlangstonsmith@smu.edu>
Date: Wed, 14 Oct 2026 15:13:44 -0500
Subject: [PATCH 02/10] Add build script for testing

---
 containers/rocker_ml_cuda/README.md          |  45 ++++++++
 containers/rocker_ml_cuda/build_container.sh | 109 +++++++++++++++++++
 containers/rocker_ml_cuda/rocker_ml_cuda.def |  22 ++++
 3 files changed, 176 insertions(+)
 create mode 100644 containers/rocker_ml_cuda/README.md
 create mode 100755 containers/rocker_ml_cuda/build_container.sh
 create mode 100644 containers/rocker_ml_cuda/rocker_ml_cuda.def

diff --git a/containers/rocker_ml_cuda/README.md b/containers/rocker_ml_cuda/README.md
new file mode 100644
index 00000000..9ef36b5b
--- /dev/null
+++ b/containers/rocker_ml_cuda/README.md
@@ -0,0 +1,45 @@
+# rocker_ml_cuda
+
+GPU-enabled RStudio container: `rocker/ml` plus CUDA PyTorch (used from R
+through reticulate), `tabpfn`, and tidymodels packages. Currently in testing
+for the OOD RStudio app.
+
+## Pins
+
+- **Base image:** `rocker/ml` is pinned by digest in `rocker_ml_cuda.def`
+  (it was `py3.14` / `latest` as of 2026-10-04). `rocker/ml` no longer
+  publishes R version tags, so a tag alone would change under us on rebuild.
+- **PyTorch:** `torch==2.14.0` from the `cu126` index. This is the last
+  release with a wheel for Volta (V100) GPUs, so don't upgrade it while
+  V100s are in use.
+- **Driver:** tested on nodes with NVIDIA driver 525.147.05 (CUDA 12.0).
+
+## Build
+
+On a cluster node with apptainer (needs network access):
+
+```
+./build_container.sh
+```
+
+This builds the image, checks that `rserver`, `rsession`, `R`, and
+`rstudio-server` exist inside it, moves it to
+`/hpc/<cluster>/containers/rocker/`, and writes
+`modules/testing/rocker_ml_cuda/<version>.lua`.
+
+## Check the GPU
+
+```
+apptainer exec --nv /hpc/<cluster>/containers/rocker/rocker_ml_cuda_1.0.0.sif nvidia-smi
+apptainer exec --nv /hpc/<cluster>/containers/rocker/rocker_ml_cuda_1.0.0.sif \
+  /opt/venv/bin/python -c 'import torch; print(torch.cuda.is_available())'
+```
+
+## Notes for OOD
+
+R is installed from apt in this image (`/usr/bin/R`, libR in
+`/usr/lib/R/lib`), not under `/usr/local` like other rocker images. The
+RStudio app's rsession wrapper must also add `/.singularity.d/libs` and
+`/opt/venv/bin` back to the paths after sourcing the host environment, or
+the GPU and the Python venv are not visible in RStudio. See
+`ood/rstudio/template/script.sh.erb`.
diff --git a/containers/rocker_ml_cuda/build_container.sh b/containers/rocker_ml_cuda/build_container.sh
new file mode 100755
index 00000000..dfc05e7a
--- /dev/null
+++ b/containers/rocker_ml_cuda/build_container.sh
@@ -0,0 +1,109 @@
+#!/bin/bash -e
+# build GPU-enabled apptainer container for RStudio (rocker/ml + CUDA PyTorch)
+# base image and package pins are in rocker_ml_cuda.def
+
+# specify version
+VERSION="1.0.0"
+IMAGE_NAME="rocker_ml_cuda"
+CONTAINER_NAME=${IMAGE_NAME}_${VERSION}.sif
+
+echo "Building ${CONTAINER_NAME}"
+
+# module file path below is relative to this directory
+cd "$(dirname "$0")"
+
+# build the container
+module purge
+module load apptainer
+unset APPTAINER_BIND
+apptainer build --fakeroot ${CONTAINER_NAME} ${IMAGE_NAME}.def
+
+# get the paths for rserver, rsession, and R inside the container. We need
+# these for the portal, so it's useful to "precompute" and load
+RSESSION_EXEC="/usr/lib/rstudio-server/bin/rsession"
+RSERVER_EXEC="/usr/lib/rstudio-server/bin/rserver"
+R_EXEC=$(apptainer exec ${CONTAINER_NAME} which R)
+# verify
+for EXEC in ${RSESSION_EXEC} ${RSERVER_EXEC} ${R_EXEC}
+do
+  if ! apptainer exec ${CONTAINER_NAME} test -x ${EXEC} &> /dev/null
+  then
+    echo "${EXEC} not found in the container"
+    exit 12
+  fi
+done
+if ! apptainer exec ${CONTAINER_NAME} which rstudio-server &> /dev/null
+then
+  echo "rstudio-server not found in the container (needed for the OOD idle timeout)"
+  exit 12
+fi
+
+# move container to /hpc/{sys}/containers/
+CLUSTER=$(scontrol show config | grep ClusterName | grep -oP '= \K.+')
+if [ "$CLUSTER" = "nvidia" ]; then
+  CLUSTER="mp"
+fi
+mkdir -p /hpc/${CLUSTER}/containers/rocker
+mv ${CONTAINER_NAME} /hpc/${CLUSTER}/containers/rocker/${CONTAINER_NAME}
+
+# create a module file in ../../modules/ (in the git repo)
+RUN_COMMAND="apptainer run --nv "
+
+mkdir -p ../../modules/testing/${IMAGE_NAME}
+MODULE_FILE=../../modules/testing/${IMAGE_NAME}/${VERSION}.lua
+(
+sed 's/^ \{2\}//' > "$MODULE_FILE" << EOL
+
+  help([[
+  Name: R + CUDA PyTorch
+  Version: ${VERSION}
+  Website: https://www.r-project.org/
+
+  R is a language and environment for statistical computing and graphics.
+  This container includes GPU-enabled PyTorch (via reticulate) for V100 GPUs.
+
+  To use R:
+
+  R -- launches an R terminal
+  Rscript script -- run a script
+
+  ]])
+  whatis("Name: R + CUDA PyTorch")
+  whatis("Version: rocker/ml:py3.14 + torch 2.14.0 (cu126)")
+  whatis("Category: R")
+  whatis("URL: https://hub.docker.com/r/rocker/ml")
+  whatis("Description: Provides access to R and Rstudio with GPU-enabled PyTorch through a Rocker container built with Apptainer")
+  family("R")
+
+  always_load('apptainer')
+  local sif_file = '/hpc/${CLUSTER}/containers/rocker/${CONTAINER_NAME}'
+
+  local work_dir = '/work'
+  local lustre_work = '/lustre/work/client'
+  local scratch_dir = os.getenv("SCRATCH")
+  local home = os.getenv("HOME")
+  local user_libs = pathJoin(home, 'R/${IMAGE_NAME}/${VERSION}')
+
+  function build_command(app)
+    local cmd        = '${RUN_COMMAND} --env R_LIBS_USER=' .. user_libs .. ' -B ' .. scratch_dir .. ',' .. work_dir .. ',' .. lustre_work .. ' ' .. sif_file .. ' ' .. app
+    local sh_ending  = ' "\$@"'
+    local csh_ending = ' \$*'
+    local sh_cmd     = cmd .. sh_ending
+    local csh_cmd    = cmd .. csh_ending
+    set_shell_function(app, sh_cmd, csh_cmd)
+  end
+
+  setenv('CONTAINER_RSESSION', '${RSESSION_EXEC}')
+  setenv('CONTAINER_RSERVER', '${RSERVER_EXEC}')
+  setenv('CONTAINER_R', '${R_EXEC}')
+  setenv('CONTAINER_IMAGE', '/hpc/${CLUSTER}/containers/rocker/${CONTAINER_NAME}')
+  setenv('R_LIBS_USER', user_libs)
+  unsetenv('XDG_RUNTIME_DIR')
+
+  build_command('R')
+  build_command('Rscript')
+EOL
+)
+
+echo "Container: /hpc/${CLUSTER}/containers/rocker/${CONTAINER_NAME}"
+echo "Module:    ${MODULE_FILE}"
diff --git a/containers/rocker_ml_cuda/rocker_ml_cuda.def b/containers/rocker_ml_cuda/rocker_ml_cuda.def
new file mode 100644
index 00000000..ada654a4
--- /dev/null
+++ b/containers/rocker_ml_cuda/rocker_ml_cuda.def
@@ -0,0 +1,22 @@
+Bootstrap: docker
+# rocker/ml:py3.14 (== latest as of 2026-10-04), pinned by digest because
+# rocker/ml no longer publishes R version tags
+From: rocker/ml@sha256:77443c5ff6422556809910000db4fb3576cb00609f39474bdb9f04edc3eeddca
+
+%labels
+    Author Mateo Langston Smith <mlangstonsmith@smu.edu>
+    Version 1.0.0
+
+%post
+    export DEBIAN_FRONTEND=noninteractive
+
+    # GPU PyTorch for V100 (cu126; 2.14.0 is the last release with a Volta wheel)
+    /opt/venv/bin/pip install --no-cache-dir torch==2.14.0 --index-url https://download.pytorch.org/whl/cu126
+    /opt/venv/bin/pip install --no-cache-dir tabpfn
+
+    # R packages (uses the image's default CRAN/binary repo settings)
+    Rscript -e 'install.packages(c("tidymodels","probably","ranger","tictoc","modeldata","Lahman","tabpfn"))'
+
+%environment
+    export CUDA_MODULE_LOADING=LAZY
+    export RETICULATE_PYTHON=/opt/venv/bin/python

From 2ffb1dfaec2e4f32157fb77593fbb63f4b93259c Mon Sep 17 00:00:00 2001
From: Mateo Langston Smith <mlangstonsmith@smu.edu>
Date: Thu, 15 Oct 2026 10:25:09 -0500
Subject: [PATCH 03/10] Move to test container

---
 ood/rstudio/form.yml.erb           | 2 +-
 ood/rstudio/template/script.sh.erb | 2 +-
 2 files changed, 2 insertions(+), 2 deletions(-)

diff --git a/ood/rstudio/form.yml.erb b/ood/rstudio/form.yml.erb
index e30f66a6..a6d02fb3 100644
--- a/ood/rstudio/form.yml.erb
+++ b/ood/rstudio/form.yml.erb
@@ -69,7 +69,7 @@ attributes:
       - [ "R 4.5.1 (custom kbnk)", "kbnk/R/4.5.1" ]
       <%- end -%>
       <%- if gpu_tester -%>
-      - [ "GPU test container (rocker-ml-cuda)", "custom" ]
+      - [ "4.6.1 + GPU", "custom" ]
       <%- end -%>
   environment:
     widget: text_area
diff --git a/ood/rstudio/template/script.sh.erb b/ood/rstudio/template/script.sh.erb
index 4e1df900..7b5e7550 100755
--- a/ood/rstudio/template/script.sh.erb
+++ b/ood/rstudio/template/script.sh.erb
@@ -15,7 +15,7 @@ env
 <%- if context.r_module == "custom" -%>
 # testing: GPU-enabled custom container, set what the R module would
 module load apptainer
-export CONTAINER_IMAGE=/lustre/scratch/client/users/mlangstonsmith/rocker-ml-cuda.sif
+export CONTAINER_IMAGE=/lustre/scratch/client/users/mlangstonsmith/rocker_ml_cuda_test.sif
 export CONTAINER_RSERVER=/usr/lib/rstudio-server/bin/rserver
 export CONTAINER_RSESSION=/usr/lib/rstudio-server/bin/rsession
 export CONTAINER_R=/usr/bin/R

From b8fbe9c8e127c9a0c0802313a07c74fe8ab359cf Mon Sep 17 00:00:00 2001
From: Mateo Langston Smith <mlangstonsmith@smu.edu>
Date: Thu, 15 Oct 2026 10:37:53 -0500
Subject: [PATCH 04/10] Add module commands

---
 ood/rstudio/form.yml.erb           |  4 ++--
 ood/rstudio/template/script.sh.erb | 14 +-------------
 2 files changed, 3 insertions(+), 15 deletions(-)

diff --git a/ood/rstudio/form.yml.erb b/ood/rstudio/form.yml.erb
index a6d02fb3..6465a52c 100644
--- a/ood/rstudio/form.yml.erb
+++ b/ood/rstudio/form.yml.erb
@@ -5,7 +5,7 @@ require 'open3'
 require_relative "CheckQueues"
 require_relative "CheckAccounts"
 
-# testing: GPU-enabled custom container (bypasses the R module)
+# testing: GPU-enabled container (modules/testing/rocker_ml_cuda)
 gpu_tester = ["mlangstonsmith"].include?(User.new.name)
 
 -%>
@@ -69,7 +69,7 @@ attributes:
       - [ "R 4.5.1 (custom kbnk)", "kbnk/R/4.5.1" ]
       <%- end -%>
       <%- if gpu_tester -%>
-      - [ "4.6.1 + GPU", "custom" ]
+      - [ "4.6.1 + GPU", "/hpc/m3/hpc_docs/modules/testing/rocker_ml_cuda/1.0.0.lua" ]
       <%- end -%>
   environment:
     widget: text_area
diff --git a/ood/rstudio/template/script.sh.erb b/ood/rstudio/template/script.sh.erb
index 7b5e7550..69f700dc 100755
--- a/ood/rstudio/template/script.sh.erb
+++ b/ood/rstudio/template/script.sh.erb
@@ -12,17 +12,6 @@ env
 <%= context.environment.encode(universal_newline: true) %>
 <%- end -%>
 
-<%- if context.r_module == "custom" -%>
-# testing: GPU-enabled custom container, set what the R module would
-module load apptainer
-export CONTAINER_IMAGE=/lustre/scratch/client/users/mlangstonsmith/rocker_ml_cuda_test.sif
-export CONTAINER_RSERVER=/usr/lib/rstudio-server/bin/rserver
-export CONTAINER_RSESSION=/usr/lib/rstudio-server/bin/rsession
-export CONTAINER_R=/usr/bin/R
-export R_LIBS_USER=${HOME}/R/rocker-ml-cuda
-mkdir -p ${R_LIBS_USER}
-unset XDG_RUNTIME_DIR
-<%- else -%>
 # load the requested module
 FORM_MODULE=<%= context.r_module %>
 EXTENSION="${FORM_MODULE##*.}"
@@ -38,7 +27,6 @@ then
 else
   module load ${FORM_MODULE}
 fi
-<%- end -%>
 
 #
 # Start RStudio Server
@@ -65,7 +53,7 @@ sed 's/^ \{2\}//' > "${RSESSION_WRAPPER_FILE}" << EOL
 
   # make sure the container libR is visible inside container
   export LD_LIBRARY_PATH="/usr/local/lib/R/lib${LD_LIBRARY_PATH:+":$LD_LIBRARY_PATH"}"
-<%- if context.r_module == "custom" -%>
+<%- if context.r_module.include?("rocker_ml_cuda") -%>
 
   # environment.txt and the line above replace the container's paths, so
   # restore libR (apt-installed in this image), the host driver libs bound

From 713e3f081cf39b786c028103167df0e12817ee3d Mon Sep 17 00:00:00 2001
From: Mateo Langston Smith <mlangstonsmith@smu.edu>
Date: Thu, 15 Oct 2026 12:28:07 -0500
Subject: [PATCH 05/10] Adjust to testing path

---
 containers/rocker_ml_cuda/README.md          |  6 +++---
 containers/rocker_ml_cuda/build_container.sh | 14 ++++++++------
 2 files changed, 11 insertions(+), 9 deletions(-)

diff --git a/containers/rocker_ml_cuda/README.md b/containers/rocker_ml_cuda/README.md
index 9ef36b5b..a354b4f6 100644
--- a/containers/rocker_ml_cuda/README.md
+++ b/containers/rocker_ml_cuda/README.md
@@ -24,14 +24,14 @@ On a cluster node with apptainer (needs network access):
 
 This builds the image, checks that `rserver`, `rsession`, `R`, and
 `rstudio-server` exist inside it, moves it to
-`/hpc/<cluster>/containers/rocker/`, and writes
+`/hpc/<cluster>/containers/testing/rocker/`, and writes
 `modules/testing/rocker_ml_cuda/<version>.lua`.
 
 ## Check the GPU
 
 ```
-apptainer exec --nv /hpc/<cluster>/containers/rocker/rocker_ml_cuda_1.0.0.sif nvidia-smi
-apptainer exec --nv /hpc/<cluster>/containers/rocker/rocker_ml_cuda_1.0.0.sif \
+apptainer exec --nv /hpc/<cluster>/containers/testing/rocker/rocker_ml_cuda_1.0.0.sif nvidia-smi
+apptainer exec --nv /hpc/<cluster>/containers/testing/rocker/rocker_ml_cuda_1.0.0.sif \
   /opt/venv/bin/python -c 'import torch; print(torch.cuda.is_available())'
 ```
 
diff --git a/containers/rocker_ml_cuda/build_container.sh b/containers/rocker_ml_cuda/build_container.sh
index dfc05e7a..2765c546 100755
--- a/containers/rocker_ml_cuda/build_container.sh
+++ b/containers/rocker_ml_cuda/build_container.sh
@@ -38,13 +38,15 @@ then
   exit 12
 fi
 
-# move container to /hpc/{sys}/containers/
+# move container to /hpc/{sys}/containers/testing/ (kept out of the main
+# containers directory while in testing)
 CLUSTER=$(scontrol show config | grep ClusterName | grep -oP '= \K.+')
 if [ "$CLUSTER" = "nvidia" ]; then
   CLUSTER="mp"
 fi
-mkdir -p /hpc/${CLUSTER}/containers/rocker
-mv ${CONTAINER_NAME} /hpc/${CLUSTER}/containers/rocker/${CONTAINER_NAME}
+CONTAINER_DIR=/hpc/${CLUSTER}/containers/testing/rocker
+mkdir -p ${CONTAINER_DIR}
+mv ${CONTAINER_NAME} ${CONTAINER_DIR}/${CONTAINER_NAME}
 
 # create a module file in ../../modules/ (in the git repo)
 RUN_COMMAND="apptainer run --nv "
@@ -76,7 +78,7 @@ sed 's/^ \{2\}//' > "$MODULE_FILE" << EOL
   family("R")
 
   always_load('apptainer')
-  local sif_file = '/hpc/${CLUSTER}/containers/rocker/${CONTAINER_NAME}'
+  local sif_file = '${CONTAINER_DIR}/${CONTAINER_NAME}'
 
   local work_dir = '/work'
   local lustre_work = '/lustre/work/client'
@@ -96,7 +98,7 @@ sed 's/^ \{2\}//' > "$MODULE_FILE" << EOL
   setenv('CONTAINER_RSESSION', '${RSESSION_EXEC}')
   setenv('CONTAINER_RSERVER', '${RSERVER_EXEC}')
   setenv('CONTAINER_R', '${R_EXEC}')
-  setenv('CONTAINER_IMAGE', '/hpc/${CLUSTER}/containers/rocker/${CONTAINER_NAME}')
+  setenv('CONTAINER_IMAGE', '${CONTAINER_DIR}/${CONTAINER_NAME}')
   setenv('R_LIBS_USER', user_libs)
   unsetenv('XDG_RUNTIME_DIR')
 
@@ -105,5 +107,5 @@ sed 's/^ \{2\}//' > "$MODULE_FILE" << EOL
 EOL
 )
 
-echo "Container: /hpc/${CLUSTER}/containers/rocker/${CONTAINER_NAME}"
+echo "Container: ${CONTAINER_DIR}/${CONTAINER_NAME}"
 echo "Module:    ${MODULE_FILE}"

From 556ebb9d902aa6a65f3091e45813c5fd3d52aa2e Mon Sep 17 00:00:00 2001
From: Mateo Langston Smith <mlangstonsmith@smu.edu>
Date: Thu, 15 Oct 2026 12:43:29 -0500
Subject: [PATCH 06/10] Add module to scratch just for testing

---
 containers/rocker_ml_cuda/README.md          |  2 +-
 containers/rocker_ml_cuda/build_container.sh | 14 ++++++++++----
 2 files changed, 11 insertions(+), 5 deletions(-)

diff --git a/containers/rocker_ml_cuda/README.md b/containers/rocker_ml_cuda/README.md
index a354b4f6..50ac5fe5 100644
--- a/containers/rocker_ml_cuda/README.md
+++ b/containers/rocker_ml_cuda/README.md
@@ -25,7 +25,7 @@ On a cluster node with apptainer (needs network access):
 This builds the image, checks that `rserver`, `rsession`, `R`, and
 `rstudio-server` exist inside it, moves it to
 `/hpc/<cluster>/containers/testing/rocker/`, and writes
-`modules/testing/rocker_ml_cuda/<version>.lua`.
+`$SCRATCH/modules/testing/rocker_ml_cuda/<version>.lua`.
 
 ## Check the GPU
 
diff --git a/containers/rocker_ml_cuda/build_container.sh b/containers/rocker_ml_cuda/build_container.sh
index 2765c546..ff737a05 100755
--- a/containers/rocker_ml_cuda/build_container.sh
+++ b/containers/rocker_ml_cuda/build_container.sh
@@ -9,9 +9,15 @@ CONTAINER_NAME=${IMAGE_NAME}_${VERSION}.sif
 
 echo "Building ${CONTAINER_NAME}"
 
-# module file path below is relative to this directory
+# the .def file path below is relative to this directory
 cd "$(dirname "$0")"
 
+# module file is written to $SCRATCH for testing
+if [ -z "${SCRATCH}" ]; then
+  echo "SCRATCH is not set"
+  exit 12
+fi
+
 # build the container
 module purge
 module load apptainer
@@ -48,11 +54,11 @@ CONTAINER_DIR=/hpc/${CLUSTER}/containers/testing/rocker
 mkdir -p ${CONTAINER_DIR}
 mv ${CONTAINER_NAME} ${CONTAINER_DIR}/${CONTAINER_NAME}
 
-# create a module file in ../../modules/ (in the git repo)
+# create a module file in $SCRATCH/modules/ (outside the git repo for testing)
 RUN_COMMAND="apptainer run --nv "
 
-mkdir -p ../../modules/testing/${IMAGE_NAME}
-MODULE_FILE=../../modules/testing/${IMAGE_NAME}/${VERSION}.lua
+mkdir -p ${SCRATCH}/modules/testing/${IMAGE_NAME}
+MODULE_FILE=${SCRATCH}/modules/testing/${IMAGE_NAME}/${VERSION}.lua
 (
 sed 's/^ \{2\}//' > "$MODULE_FILE" << EOL
 

From 1464bf9dcd89ed7df855f5fd9c00e74c00cec2ae Mon Sep 17 00:00:00 2001
From: Mateo Langston Smith <mlangstonsmith@smu.edu>
Date: Thu, 15 Oct 2026 12:47:08 -0500
Subject: [PATCH 07/10] relative path

---
 containers/rocker_ml_cuda/README.md          |  2 +-
 containers/rocker_ml_cuda/build_container.sh | 14 ++++----------
 2 files changed, 5 insertions(+), 11 deletions(-)

diff --git a/containers/rocker_ml_cuda/README.md b/containers/rocker_ml_cuda/README.md
index 50ac5fe5..c6b5cf0c 100644
--- a/containers/rocker_ml_cuda/README.md
+++ b/containers/rocker_ml_cuda/README.md
@@ -25,7 +25,7 @@ On a cluster node with apptainer (needs network access):
 This builds the image, checks that `rserver`, `rsession`, `R`, and
 `rstudio-server` exist inside it, moves it to
 `/hpc/<cluster>/containers/testing/rocker/`, and writes
-`$SCRATCH/modules/testing/rocker_ml_cuda/<version>.lua`.
+`modules/testing/rocker_ml_cuda/<version>.lua` in this repo.
 
 ## Check the GPU
 
diff --git a/containers/rocker_ml_cuda/build_container.sh b/containers/rocker_ml_cuda/build_container.sh
index ff737a05..414ac25f 100755
--- a/containers/rocker_ml_cuda/build_container.sh
+++ b/containers/rocker_ml_cuda/build_container.sh
@@ -9,15 +9,9 @@ CONTAINER_NAME=${IMAGE_NAME}_${VERSION}.sif
 
 echo "Building ${CONTAINER_NAME}"
 
-# the .def file path below is relative to this directory
+# .def and module file paths below are relative to this directory
 cd "$(dirname "$0")"
 
-# module file is written to $SCRATCH for testing
-if [ -z "${SCRATCH}" ]; then
-  echo "SCRATCH is not set"
-  exit 12
-fi
-
 # build the container
 module purge
 module load apptainer
@@ -54,11 +48,11 @@ CONTAINER_DIR=/hpc/${CLUSTER}/containers/testing/rocker
 mkdir -p ${CONTAINER_DIR}
 mv ${CONTAINER_NAME} ${CONTAINER_DIR}/${CONTAINER_NAME}
 
-# create a module file in $SCRATCH/modules/ (outside the git repo for testing)
+# create a module file in ../../modules/ (in the git repo)
 RUN_COMMAND="apptainer run --nv "
 
-mkdir -p ${SCRATCH}/modules/testing/${IMAGE_NAME}
-MODULE_FILE=${SCRATCH}/modules/testing/${IMAGE_NAME}/${VERSION}.lua
+mkdir -p ../../modules/testing/${IMAGE_NAME}
+MODULE_FILE=../../modules/testing/${IMAGE_NAME}/${VERSION}.lua
 (
 sed 's/^ \{2\}//' > "$MODULE_FILE" << EOL
 

From e7345a7045aa52ce1360b44bdd07873c0109e101 Mon Sep 17 00:00:00 2001
From: mateolangston <mlangstonsmith@smu.edu>
Date: Thu, 15 Oct 2026 13:03:40 -0500
Subject: [PATCH 08/10] add module file

---
 modules/testing/rocker_ml_cuda/1.0.0.lua | 49 ++++++++++++++++++++++++
 1 file changed, 49 insertions(+)
 create mode 100644 modules/testing/rocker_ml_cuda/1.0.0.lua

diff --git a/modules/testing/rocker_ml_cuda/1.0.0.lua b/modules/testing/rocker_ml_cuda/1.0.0.lua
new file mode 100644
index 00000000..c6d8bace
--- /dev/null
+++ b/modules/testing/rocker_ml_cuda/1.0.0.lua
@@ -0,0 +1,49 @@
+
+help([[
+Name: R + CUDA PyTorch
+Version: 1.0.0
+Website: https://www.r-project.org/
+
+R is a language and environment for statistical computing and graphics.
+This container includes GPU-enabled PyTorch (via reticulate) for V100 GPUs.
+
+To use R:
+
+R -- launches an R terminal
+Rscript script -- run a script
+
+]])
+whatis("Name: R + CUDA PyTorch")
+whatis("Version: rocker/ml:py3.14 + torch 2.14.0 (cu126)")
+whatis("Category: R")
+whatis("URL: https://hub.docker.com/r/rocker/ml")
+whatis("Description: Provides access to R and Rstudio with GPU-enabled PyTorch through a Rocker container built with Apptainer")
+family("R")
+
+always_load('apptainer')
+local sif_file = '/hpc/m3/containers/testing/rocker/rocker_ml_cuda_1.0.0.sif'
+
+local work_dir = '/work'
+local lustre_work = '/lustre/work/client'
+local scratch_dir = os.getenv("SCRATCH")
+local home = os.getenv("HOME")
+local user_libs = pathJoin(home, 'R/rocker_ml_cuda/1.0.0')
+
+function build_command(app)
+  local cmd        = 'apptainer run --nv  --env R_LIBS_USER=' .. user_libs .. ' -B ' .. scratch_dir .. ',' .. work_dir .. ',' .. lustre_work .. ' ' .. sif_file .. ' ' .. app
+  local sh_ending  = ' "$@"'
+  local csh_ending = ' $*'
+  local sh_cmd     = cmd .. sh_ending
+  local csh_cmd    = cmd .. csh_ending
+  set_shell_function(app, sh_cmd, csh_cmd)
+end
+
+setenv('CONTAINER_RSESSION', '/usr/lib/rstudio-server/bin/rsession')
+setenv('CONTAINER_RSERVER', '/usr/lib/rstudio-server/bin/rserver')
+setenv('CONTAINER_R', '/usr/bin/R')
+setenv('CONTAINER_IMAGE', '/hpc/m3/containers/testing/rocker/rocker_ml_cuda_1.0.0.sif')
+setenv('R_LIBS_USER', user_libs)
+unsetenv('XDG_RUNTIME_DIR')
+
+build_command('R')
+build_command('Rscript')

From 31aa918e4d4a3f57f91fba217719859e28189b17 Mon Sep 17 00:00:00 2001
From: Mateo Langston Smith <mlangstonsmith@smu.edu>
Date: Thu, 15 Oct 2026 13:09:46 -0500
Subject: [PATCH 09/10] Update form.yml.erb

---
 ood/rstudio/form.yml.erb | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/ood/rstudio/form.yml.erb b/ood/rstudio/form.yml.erb
index 6465a52c..12f21092 100644
--- a/ood/rstudio/form.yml.erb
+++ b/ood/rstudio/form.yml.erb
@@ -69,7 +69,8 @@ attributes:
       - [ "R 4.5.1 (custom kbnk)", "kbnk/R/4.5.1" ]
       <%- end -%>
       <%- if gpu_tester -%>
-      - [ "4.6.1 + GPU", "/hpc/m3/hpc_docs/modules/testing/rocker_ml_cuda/1.0.0.lua" ]
+      # TODO before PR: change base path back to /hpc/m3/hpc_docs/modules
+      - [ "4.6.1 + GPU", "/lustre/scratch/client/users/mlangstonsmith/hpc_docs/modules/testing/rocker_ml_cuda/1.0.0.lua" ]
       <%- end -%>
   environment:
     widget: text_area

From c9f809bfb2a8b80f458eb48e0505a9bb28fd065f Mon Sep 17 00:00:00 2001
From: Mateo Langston Smith <mlangstonsmith@smu.edu>
Date: Thu, 15 Oct 2026 13:50:42 -0500
Subject: [PATCH 10/10] Update form.yml.erb

---
 ood/rstudio/form.yml.erb | 10 +++-------
 1 file changed, 3 insertions(+), 7 deletions(-)

diff --git a/ood/rstudio/form.yml.erb b/ood/rstudio/form.yml.erb
index 12f21092..1eef12af 100644
--- a/ood/rstudio/form.yml.erb
+++ b/ood/rstudio/form.yml.erb
@@ -5,9 +5,6 @@ require 'open3'
 require_relative "CheckQueues"
 require_relative "CheckAccounts"
 
-# testing: GPU-enabled container (modules/testing/rocker_ml_cuda)
-gpu_tester = ["mlangstonsmith"].include?(User.new.name)
-
 -%>
 
 ---
@@ -68,9 +65,8 @@ attributes:
       <%- if CheckAccounts.Accounts.any? { |word| word.include?("kbnk") } | CheckAccounts.Accounts.any? { |word| word.include?("jlagrone") }-%>
       - [ "R 4.5.1 (custom kbnk)", "kbnk/R/4.5.1" ]
       <%- end -%>
-      <%- if gpu_tester -%>
-      # TODO before PR: change base path back to /hpc/m3/hpc_docs/modules
-      - [ "4.6.1 + GPU", "/lustre/scratch/client/users/mlangstonsmith/hpc_docs/modules/testing/rocker_ml_cuda/1.0.0.lua" ]
+      <%- if CheckAccounts.Accounts.any? { |word| word.include?("csouth") } | CheckAccounts.Accounts.any? { |word| word.include?("jlagrone") }-%>
+      - [ "4.6.1 + GPU", "/hpc/m3/hpc_docs/modules/testing/rocker_ml_cuda/1.0.0.lua" ]
       <%- end -%>
   environment:
     widget: text_area
@@ -87,7 +83,7 @@ attributes:
       [SLURM partition]: https://southernmethodistuniversity.github.io/hpc_docs/about/queues.html#m3-queues
     options:
     <%- CheckQueues.Queues.each do |q| -%>
-      <%- if q[4].to_i == 0 || gpu_tester -%> #disable gpu queues (except for testers)
+      <%- if q[4].to_i == 0 || CheckAccounts.Accounts.any? { |word| word.include?("csouth") } || CheckAccounts.Accounts.any? { |word| word.include?("jlagrone") } -%> #disable gpu queues (except for GPU testers)
       - ["<%= q[0] %>", "<%= q[0] %>",
         data-max-bc-num-hours: "<%= q[1] %>",
         data-max-bc-num-cores: "<%= q[2] %>",