fix(neuron): install aws-neuronx-dkms-2.21 at boot on inf1

Author: mselim00

templates/al2023/provisioners/install-neuron-driver.sh

@@ -4,6 +4,8 @@ set -o pipefail
 set -o nounset
 set -o errexit
 
+readonly PACKAGE_CACHE_PATH="/var/cache/eks/packages"
+
 if [ "$ENABLE_ACCELERATOR" != "neuron" ]; then
   exit 0
 fi
@@ -24,6 +26,27 @@ EOF
 # Manually install the GPG key, verifies repository can be reached
 sudo rpm --import https://yum.repos.neuron.amazonaws.com/GPG-PUB-KEY-AMAZON-AWS-NEURON.PUB
 
+################################################################################
+### Cache packages for conditional install at boot #############################
+################################################################################
+# TODO: remove this section if inf1 is not supported
+
+# Install and remove aws-neuronx-dkms-2.21.x to ensure all of its dependencies are
+# pre-installed
+sudo dnf install aws-neuronx-dkms-2.21.*
+sudo dnf remove --noautoremove aws-neuronx-dkms
+
+# Cache the 2.21.x rpm for contidtional boot-time install
+sudo dnf download aws-neuronx-dkms-2.21.*
+sudo mkdir -p "$PACKAGE_CACHE_PATH"
+sudo mv aws-neuronx-dkms-*.rpm "${PACKAGE_CACHE_PATH}/"
+
+sudo mv ${WORKING_DIR}/gpu/neuron-package-install.sh /etc/eks/
+sudo mv ${WORKING_DIR}/gpu/neuron-package-install.service /etc/systemd/system/
+
+sudo systemctl daemon-reload
+sudo systemctl enable neuron-package-install.service
+
 ################################################################################
 ### Install packages ###########################################################
 ################################################################################

templates/al2023/runtime/gpu/neuron-package-install.service

@@ -0,0 +1,13 @@
+[Unit]
+Description=Install Neuron packages
+# Run before cloud-init so packages are installed
+# before user data that may query the installed information
+Before=cloud-init.service
+
+[Service]
+Type=oneshot
+ExecStart=/etc/eks/neuron-package-install.sh
+RemainAfterExit=true
+
+[Install]
+WantedBy=multi-user.target

templates/al2023/runtime/gpu/neuron-package-install.sh

@@ -0,0 +1,39 @@
+#!/usr/bin/env bash
+
+set -o errexit
+set -o pipefail
+set -o nounset
+
+readonly PACKAGE_CACHE_PATH="/var/cache/eks/packages"
+
+readonly AMAZON_VENDOR_CODE="1d0f"
+# Based on https://github.com/aws-neuron/aws-neuron-driver/blob/fca19f7df31b44cbbffaf121230a66df6f59d118/neuron_device.h#L36-L39
+readonly INF1_DEVICE_IDS=("7064" "7065" "7066" "7067")
+
+function is-inf1() {
+  for DEVICE_ID in "${INF1_DEVICE_IDS[@]}"; do
+    MATCHED_DEVICES=$(lspci -d "${AMAZON_VENDOR_CODE}:${DEVICE_ID}" | wc -l)
+    if [[ "$MATCHED_DEVICES" -gt 0 ]]; then
+      return 0
+    fi
+  done
+
+  return 1
+}
+
+function installed-neuron-driver-version() {
+  rpm -q aws-neuronx-dkms --queryformat '%{VERSION}'
+}
+
+if is-inf1 && [[ $(installed-neuron-driver-version) != 2.21.* ]]; then
+  echo "downgrading driver to 2.21"
+  # "dnf downgrade" would fail because the post remove script for the package
+  # does not fully remove the module, and then the post install script for the
+  # downgraded version fails because of an attempt to probe an older version of
+  # a loaded module without --force. relying on the rpm cli directly makes the
+  # operations more intuitive
+  sudo rpm --erase aws-neuronx-dkms
+  sudo rpm -i "${PACKAGE_CACHE_PATH}/aws-neuronx-dkms-2.21.*.rpm"
+else
+  echo "nothing to do!"
+fi