Introducing AL2023 GPU AMIs

Makefile

@@ -101,6 +101,10 @@ al2023arm: check-region init validate release-al2023.auto.pkrvars.hcl
 al2023neu: check-region init validate release-al2023.auto.pkrvars.hcl
 	./packer build -only="amazon-ebs.al2023neu" -var "region=${REGION}" .
 
+.PHONY: al2023gpu
+al2023gpu: check-region init validate release-al2023.auto.pkrvars.hcl
+	./packer build -only="amazon-ebs.al2023gpu" -var "region=${REGION}" .
+
 shellcheck:
 	curl -fLSs ${SHELLCHECK_URL} -o /tmp/shellcheck.tar.xz
 	tar -xvf /tmp/shellcheck.tar.xz -C /tmp --strip-components=1

README.md

@@ -7,7 +7,7 @@ It will create a private AMI in whatever account you are running it in.
 
 1. Setup AWS cli credentials.
 2. Make the recipe that you want, REGION must be specified. Options are: al1, al2, al2arm, al2gpu, al2keplergpu, al2inf, 
-al2kernel5dot10, al2kernel5dot10arm, al2kernel5dot10gpu, al2kernel5dot10inf, al2023, al2023arm, al2023neu.
+al2kernel5dot10, al2kernel5dot10arm, al2kernel5dot10gpu, al2kernel5dot10inf, al2023, al2023arm, al2023neu, al2023gpu.
 ```
 REGION=us-west-2 make al2
 ```

al2023.pkr.hcl

@@ -43,7 +43,8 @@ build {
   sources = [
     "source.amazon-ebs.al2023",
     "source.amazon-ebs.al2023arm",
-    "source.amazon-ebs.al2023neu"
+    "source.amazon-ebs.al2023neu",
+    "source.amazon-ebs.al2023gpu"
   ]
 
   provisioner "file" {
@@ -173,6 +174,13 @@ build {
     script              = "scripts/enable-ecs-agent-inferentia-support.sh"
   }
 
+  provisioner "shell" {
+    environment_vars = [
+      "AMI_TYPE=${source.name}"
+    ]
+    script = "scripts/enable-ecs-agent-gpu-support-al2023.sh"
+  }
+
   provisioner "shell" {
     inline_shebang = "/bin/sh -ex"
     inline = [

al2023gpu.pkr.hcl

@@ -0,0 +1,40 @@
+locals {
+  ami_name_al2023gpu = "${var.ami_name_prefix_al2023}-gpu-hvm-2023.0.${var.ami_version_al2023}${var.kernel_version_al2023}-x86_64-ebs"
+  default_tags = {
+    os_version          = "Amazon Linux 2023"
+    source_image_name   = "{{ .SourceAMIName }}"
+    ecs_runtime_version = "Docker version ${var.docker_version_al2023}"
+    ecs_agent_version   = "${var.ecs_agent_version}"
+    ami_type            = "al2023gpu"
+    ami_version         = "2023.0.${var.ami_version_al2023}"
+  }
+  merged_tags = merge("${local.default_tags}", "${var.tags}")
+}
+
+source "amazon-ebs" "al2023gpu" {
+  ami_name        = "${local.ami_name_al2023gpu}"
+  ami_description = "Amazon Linux AMI 2023.0.${var.ami_version_al2023} x86_64 ECS HVM EBS"
+  instance_type   = var.gpu_instance_types[0]
+  launch_block_device_mappings {
+    volume_size           = var.block_device_size_gb
+    delete_on_termination = true
+    volume_type           = "gp3"
+    device_name           = "/dev/xvda"
+  }
+  region = var.region
+  source_ami_filter {
+    filters = {
+      name = "${var.source_ami_al2023}"
+    }
+    owners             = ["amazon"]
+    most_recent        = true
+    include_deprecated = true
+  }
+  ami_ou_arns   = "${var.ami_ou_arns}"
+  ami_org_arns  = "${var.ami_org_arns}"
+  ami_users     = "${var.ami_users}"
+  ssh_interface = "public_ip"
+  ssh_username  = "ec2-user"
+  tags          = "${local.merged_tags}"
+  run_tags      = "${var.run_tags}"
+}

scripts/enable-ecs-agent-gpu-support-al2023.sh

@@ -0,0 +1,36 @@
+#!/usr/bin/env bash
+set -ex
+
+if [[ $AMI_TYPE != "al2023"*"gpu" ]]; then
+    exit 0
+fi
+
+# AL2023 GPU setup
+sudo dnf install -y dkms kernel-modules-extra
+sudo systemctl enable --now dkms
+
+# nvidia-release creates an nvidia repo file at /etc/yum.repos.d/amazonlinux-nvidia.repo
+# docker-runtime-nvidia is a custom Amazon Linux package that sets the gpu container runtime to nvidia
+# instead of having to use 'nvidia-ctk runtime configure --runtime=docker'
+sudo dnf install -y nvidia-release
+sudo dnf install -y nvidia-driver \
+    nvidia-fabric-manager \
+    pciutils \
+    xorg-x11-server-Xorg \
+    nvidia-container-toolkit \
+    docker-runtime-nvidia
+
+sudo dnf install -y cuda
+
+# The Fabric Manager service needs to be started and enabled on EC2 P4d instances
+# in order to configure NVLinks and NVSwitches
+sudo systemctl enable nvidia-fabricmanager
+
+# NVIDIA Persistence Daemon needs to be started and enabled on P5 instances
+# to maintain persistent software state in the NVIDIA driver.
+sudo systemctl enable nvidia-persistenced
+
+# Enable GPU support for ECS
+mkdir -p /tmp/ecs
+echo 'ECS_ENABLE_GPU_SUPPORT=true' >/tmp/ecs/ecs.config
+sudo mv /tmp/ecs/ecs.config /var/lib/ecs/ecs.config