[WIP] Compiled pytorch within MOOSE-dev container.

apptainer/moose-dev.def

@@ -29,8 +29,8 @@
 {%- set GPERF_DIR = '/opt/gperftools' -%}
 {#- The installation location for pprof                                                       -#}
 {%- set PPROF_DIR = '/opt/pprof' -%}
-{#- The installation location for libtorch                                                    -#}
-{%- set LIBTORCH_DEST = '/opt' -%}
+{#- We need this here because more than one block will use it                                 -#}
+{%- set MOOSE_PYTHON_VERSION = '3.11' -%}
 
 {#- The script used to install wasp                                                           -#}
 {%- set WASP_BUILD_SCRIPT = 'update_and_rebuild_wasp.sh' -%}
@@ -52,7 +52,7 @@ Fingerprints: 0CFFCAB55E806363601C442D211817B01E0911DB
 
 {%- if WITH_LIBTORCH %}
     # Make libtorch visible to moose
-    export LIBTORCH_DIR={{ LIBTORCH_DEST }}/libtorch
+    export LIBTORCH_DIR={{ MOOSE_MINIFORGE }}/lib/python{{ MOOSE_PYTHON_VERSION }}/site-packages/torch
     # Adding this to not get GPU initialization errors from MPICH
     export MPIR_CVAR_ENABLE_GPU=0
 {%- endif %}
@@ -81,7 +81,8 @@ Fingerprints: 0CFFCAB55E806363601C442D211817B01E0911DB
 
     # Pinned versions
     MINIFORGE_VERSION=23.3.1-1
-    PYTHON_VERSION=3.11
+    PYTHON_VERSION={{ MOOSE_PYTHON_VERSION }}
+    echo $PYTHON_VERSION
     CODE_SERVER_VERSION=4.92.2
 
     # Install code-server
@@ -113,26 +114,9 @@ Fingerprints: 0CFFCAB55E806363601C442D211817B01E0911DB
     git checkout ${WASP_GIT_SHA}
     git submodule update --init --recursive
 
-{%- if WITH_LIBTORCH %}
-    # Libtorch-related vars
-    LIBTORCH_DEST={{ LIBTORCH_DEST }}
-    LIBTORCH_VER=2.1.0
-    LIBTORCH_DISTRIBUTION={{ WITH_LIBTORCH }}
-
-    # Install libtorch from a precompiled package
-    ${ROOT_BUILD_DIR}/setup_libtorch.sh --version=${LIBTORCH_VER} --libtorch-dest=${LIBTORCH_DEST} --libtorch-distribution=${LIBTORCH_DISTRIBUTION} --cleanup
-
-{%- if WITH_LIBTORCH.startswith('cu') %}
-    # We install CUDA Toolkit if the user wants cuda-based libtorch.
-    # Right now this assumes that cuda-based distributions start with -cu-
-        dnf -y install epel-release
-        CUDA_RPM=${BUILD_DIR}/cuda.rpm
-        curl -L https://developer.download.nvidia.com/compute/cuda/11.4.0/local_installers/cuda-repo-rhel8-11-4-local-11.4.0_470.42.01-1.x86_64.rpm -o ${CUDA_RPM}
-        rpm -i ${CUDA_RPM}
-        dnf -y install cuda
-        rm -rf ${CUDA_RPM}
-{%- endif %}
-{%- endif %}
+    # Build and install wasp
+    # This is redundant; hopefully we can use the one from conda in the future
+    WASP_SRC_DIR=${WASP_SRC_DIR} ${WASP_BUILD_SCRIPT} -D CMAKE_INSTALL_PREFIX:STRING=${WASP_DIR}
 
     # Install miniforge in MOOSE_MINIFORGE
     MINIFORGE_SCRIPT=${ROOT_BUILD_DIR}/miniforge3.sh
@@ -156,9 +140,47 @@ Fingerprints: 0CFFCAB55E806363601C442D211817B01E0911DB
     # Needed for coverage
     pip3 --no-cache install lcov-cobertura
 
-    # Build and install wasp
-    # This is redundant; hopefully we can use the one from conda in the future
-    WASP_SRC_DIR=${WASP_SRC_DIR} ${WASP_BUILD_SCRIPT} -D CMAKE_INSTALL_PREFIX:STRING=${WASP_DIR}
+{%- if WITH_LIBTORCH %}
+    # Libtorch-related vars
+    LIBTORCH_VER=2.3.0
+    LIBTORCH_DISTRIBUTION={{ WITH_LIBTORCH }}
+
+    # Clone pytorch
+    cd ${ROOT_BUILD_DIR}
+    git clone -b v${LIBTORCH_VER} --recursive https://github.com/pytorch/pytorch
+    cd pytorch
+
+    # Install the python requirements
+    pip install -r requirements.txt
+
+    # Below a certain version we need to downgrade our numpy
+    if { echo ${LIBTORCH_VER}; echo "2.2.0"; } | sort --version-sort --check=quiet; then
+      conda install -yq "numpy<2.0"
+    fi
+
+    # Fetch the patch that enables the successful compilation on this linux distro
+    curl -L https://github.com/pytorch/pytorch/pull/99468.patch > 99468.patch
+    git apply 99468.patch
+
+    # We have to diasble some errors considering they emerge as some optimization of the compiler on certain distributuins (sadly on this one to)
+    export CFLAGS+=" -Wno-error=maybe-uninitialized -Wno-error=uninitialized -Wno-error=restrict"
+    export CXXFLAGS+=" -Wno-error=maybe-uninitialized -Wno-error=uninitialized -Wno-error=restrict"
+
+    # Build and install
+    export CMAKE_PREFIX_PATH="${CONDA_PREFIX:-'$(dirname $(which conda))/../'}:${CMAKE_PREFIX_PATH}"
+    MAX_JOBS=${MOOSE_JOBS} python setup.py install
+
+{%- if WITH_LIBTORCH.startswith('cu') %}
+    # We install CUDA Toolkit if the user wants cuda-based libtorch.
+    # Right now this assumes that cuda-based distributions start with -cu-
+        dnf -y install epel-release
+        CUDA_RPM=${BUILD_DIR}/cuda.rpm
+        curl -L https://developer.download.nvidia.com/compute/cuda/11.4.0/local_installers/cuda-repo-rhel8-11-4-local-11.4.0_470.42.01-1.x86_64.rpm -o ${CUDA_RPM}
+        rpm -i ${CUDA_RPM}
+        dnf -y install cuda
+        rm -rf ${CUDA_RPM}
+{%- endif %}
+{%- endif %}
 
     # Install node.js. Requested by dschwen for bison
     dnf install -y nodejs npm
@@ -206,6 +228,3 @@ Fingerprints: 0CFFCAB55E806363601C442D211817B01E0911DB
     {{ FILES_DIR }}/opt/code-server/bin/code-server-start /opt/code-server/bin/code-server-start
     {{ MOOSE_DIR }}/scripts/{{ WASP_BUILD_SCRIPT }} {{ ROOT_BUILD_DIR }}/{{ WASP_BUILD_SCRIPT }}
     {{ MOOSE_DIR }}/scripts/configure_wasp.sh {{ ROOT_BUILD_DIR }}/configure_wasp.sh
-{%- if WITH_LIBTORCH %}
-    {{ MOOSE_DIR }}/scripts/setup_libtorch.sh {{ ROOT_BUILD_DIR }}/setup_libtorch.sh
-{%- endif %}

scripts/setup_libtorch.sh

@@ -80,7 +80,7 @@ version_check() {
       loc_2=$1
       ;;
     *)
-      echo "'version_check' function does only supports '-g' and '-l' for comparison!"
+      echo "'version_check' function does only support '-g' and '-l' for comparison!"
       exit 0
       ;;
   esac