Provision AWS Infrastructure with Terraform

/opt/lilt/
├── install-lilt-eks.sh
├── terraform-modules/
│   └── lilt-aws-env/
└── terraform/
    └── main.tf          # you create this

module "lilt_env" {
  source = "../terraform-modules/lilt-aws-env"

  # Identity. Every resource name derives from prefix.
  prefix      = "acme-prod"
  subdomain   = "prod"
  base_domain = "example.com"

  # Networking, from your own VPC layer.
  vpc_id         = "vpc-0abc123def456789a"
  vpc_cidr       = "10.20.0.0/16"
  nat_gateway_id = "nat-0abc123def456789a"
  az_a           = "us-east-1a"
  az_b           = "us-east-1b"
  module_cidr_a  = "10.20.80.0/22"
  module_cidr_b  = "10.20.84.0/22"
  shared_cidrs   = ["10.20.0.0/24"]

  # Who administers the cluster. Omitting yourself locks you out.
  eks_cluster_admin_arns = {
    "operators" = "arn:aws:iam::111122223333:role/lilt-operators"
  }

  # DNS.
  r53_hosted_zone_id = "Z0EXAMPLE1234567890"

  # Database. Set rds_multi_az to true for production.
  rds_instance_class    = "db.r6g.large"
  rds_allocated_storage = 100
  rds_multi_az          = true

  # Node sizing.
  eks_application_instance_types = ["r6i.2xlarge", "r5.2xlarge"]
  eks_application_min_size       = 2
  eks_application_max_size       = 8
  eks_application_volume_size    = 300

  # Run the installer from inside the VPC. The EKS endpoint is private.
  enable_bastion = true
}

cd terraform
terraform init
terraform plan
terraform apply

locals {
  # Every zone in the region, so a GPU pool has somewhere else to go when one
  # zone is out of capacity.
  gpu_azs = ["us-east-1a", "us-east-1b", "us-east-1c", "us-east-1d", "us-east-1f"]

  gpu_labels = {
    "node-type"              = "gpu"
    "capability"             = "gpu"
    "nvidia.com/gpu.present" = "true"
  }

  gpu_taints = [{ key = "nvidia.com/gpu", value = "true", effect = "NO_SCHEDULE" }]
}

eks_extra_node_groups = {
  # 4x T4, 48 vCPU, 192 GiB - AI Review and LILT Create.
  "gpu-multi" = {
    instance_types     = ["g4dn.12xlarge"]
    min_size           = 1
    max_size           = 2
    volume_size        = 2000
    gpu                = true
    cluster_autoscaler = true
    availability_zones = local.gpu_azs
    labels             = merge(local.gpu_labels, { "gpu-class" = "multi" })
    taints             = local.gpu_taints
  }

  # 1x L40S, 48 GiB VRAM - OCR and the larger single-GPU models.
  "gpu-vram" = {
    instance_types     = ["g6e.4xlarge"]
    min_size           = 1
    max_size           = 2
    volume_size        = 2000
    gpu                = true
    cluster_autoscaler = true
    availability_zones = local.gpu_azs
    labels             = merge(local.gpu_labels, { "gpu-class" = "vram" })
    taints             = local.gpu_taints
  }

  # 4x L4, 48 vCPU, 192 GiB - translation, vmf and the batch workers.
  "gpu-standard" = {
    instance_types     = ["g6.12xlarge"]
    min_size           = 1
    max_size           = 2
    volume_size        = 2000
    gpu                = true
    cluster_autoscaler = true
    availability_zones = local.gpu_azs
    labels             = merge(local.gpu_labels, { "gpu-class" = "standard" })
    taints             = local.gpu_taints
  }

  # 1x T4, 8 vCPU, 32 GiB - speech recognition.
  "gpu-whisper" = {
    instance_types     = ["g4dn.2xlarge"]
    min_size           = 1
    max_size           = 2
    volume_size        = 1000
    gpu                = true
    cluster_autoscaler = true
    availability_zones = local.gpu_azs
    labels             = merge(local.gpu_labels, { "gpu-class" = "whisper" })
    taints             = local.gpu_taints
  }
}

terraform output -json