@aws/ml-container-creator 0.9.0 → 0.10.0

package/templates/do/deploy

@@ -1,1766 +1 @@
-#!/bin/bash
-# Copyright Amazon.com, Inc. or its affiliates. All Rights Reserved.
-# SPDX-License-Identifier: Apache-2.0
-
-set -e
-set -u
-set -o pipefail
-
-# Parse flags
-FORCE_NEW=false
-FORCE_IC=false
-IC_TARGET=""
-while [ $# -gt 0 ]; do
-    case "$1" in
-        --force) FORCE_NEW=true; shift ;;
-        --force-ic)
-            FORCE_IC=true
-            shift
-<% if (deploymentTarget === 'realtime-inference') { %>
-            # Optional name argument: --force-ic <name>
-            if [ $# -gt 0 ] && [[ ! "$1" == --* ]]; then
-                IC_TARGET="$1"
-                shift
-            fi
-<% } %>
-            ;;
-<% if (deploymentTarget === 'realtime-inference') { %>
-        --ic)
-            if [ -z "${2:-}" ]; then
-                echo "❌ --ic requires a name argument"
-                echo "   Usage: ./do/deploy --ic <name>"
-                exit 1
-            fi
-            IC_TARGET="$2"
-            shift 2
-            ;;
-<% } %>
-        --help|-h)
-<% if (deploymentTarget === 'realtime-inference') { %>
-            echo "Usage: ./do/deploy [--force] [--force-ic [<name>]] [--ic <name>]"
-            echo ""
-            echo "Options:"
-            echo "  --force            Create a new endpoint and IC, even if one already exists."
-            echo "  --force-ic         Recreate ALL inference components on the existing endpoint."
-            echo "  --force-ic <name>  Recreate only the named IC on the existing endpoint."
-            echo "  --ic <name>        Deploy only the named IC (from do/ic/<name>.conf)."
-            echo ""
-            echo "Without flags, deploy resumes from the last run."
-<% } else { %>
-            echo "Usage: ./do/deploy [--force] [--force-ic]"
-            echo ""
-            echo "Options:"
-            echo "  --force      Create a new endpoint, even if one already exists."
-            echo "  --force-ic   Recreate the inference component on the existing endpoint."
-            echo ""
-            echo "Without flags, deploy resumes from the last run."
-<% } %>
-            exit 0
-            ;;
-        *)
-            echo "❌ Unknown option: $1"
-            echo "   Run ./do/deploy --help for usage."
-            exit 1
-            ;;
-    esac
-done
-
-# Source configuration
-SCRIPT_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
-source "${SCRIPT_DIR}/config"
-
-echo "🚀 Deploying to AWS"
-echo "   Project: ${PROJECT_NAME}"
-echo "   Deployment config: ${DEPLOYMENT_CONFIG}"
-echo "   Region: ${AWS_REGION}"
-echo "   Build target: ${BUILD_TARGET}"
-echo "   Deployment target: ${DEPLOYMENT_TARGET}"
-<% if (deploymentTarget === 'realtime-inference') { %>
-if [ "${ENDPOINT_EXTERNAL:-false}" = "true" ]; then
-    echo "   Endpoint: ${ENDPOINT_NAME} (external)"
-else
-    echo "   Instance type: ${INSTANCE_TYPE}"
-fi
-<% } else if (deploymentTarget === 'async-inference') { %>
-echo "   Instance type: ${INSTANCE_TYPE}"
-echo "   S3 output: ${ASYNC_S3_OUTPUT_PATH}"
-echo "   SNS success: ${ASYNC_SNS_SUCCESS_TOPIC}"
-echo "   SNS error: ${ASYNC_SNS_ERROR_TOPIC}"
-<% if (asyncMaxConcurrentInvocations) { %>
-echo "   Max concurrent: ${ASYNC_MAX_CONCURRENT_INVOCATIONS}"
-<% } %>
-<% } else if (deploymentTarget === 'hyperpod-eks') { %>
-echo "   HyperPod cluster: ${HYPERPOD_CLUSTER_NAME}"
-echo "   Namespace: ${HYPERPOD_NAMESPACE}"
-echo "   Replicas: ${HYPERPOD_REPLICAS}"
-<% } else if (deploymentTarget === 'batch-transform') { %>
-echo "   Instance type: ${INSTANCE_TYPE}"
-echo "   S3 input: ${BATCH_INPUT_PATH}"
-echo "   S3 output: ${BATCH_OUTPUT_PATH}"
-echo "   Instance count: ${BATCH_INSTANCE_COUNT}"
-echo "   Split type: ${BATCH_SPLIT_TYPE}"
-echo "   Strategy: ${BATCH_STRATEGY}"
-<% } %>
-
-# Check AWS credentials
-echo "🔍 Validating AWS credentials..."
-if ! aws sts get-caller-identity &> /dev/null; then
-    echo "❌ AWS credentials not configured"
-    echo "   Run: aws configure"
-    echo "   Or set AWS_ACCESS_KEY_ID and AWS_SECRET_ACCESS_KEY environment variables"
-    exit 4
-fi
-
-AWS_ACCOUNT_ID=$(aws sts get-caller-identity --query Account --output text)
-echo "✅ AWS credentials validated (Account: ${AWS_ACCOUNT_ID})"
-
-# Construct ECR repository URL
-ECR_REPOSITORY="${AWS_ACCOUNT_ID}.dkr.ecr.${AWS_REGION}.amazonaws.com/${ECR_REPOSITORY_NAME}"
-
-# ============================================================
-# Shared: Verify ECR image exists
-# ============================================================
-echo "🔍 Verifying ECR image exists..."
-if ! aws ecr describe-images \
-    --repository-name "${ECR_REPOSITORY_NAME}" \
-    --image-ids imageTag="${PROJECT_NAME}-latest" \
-    --region "${AWS_REGION}" &> /dev/null; then
-    
-    echo "❌ ECR image not found: ${ECR_REPOSITORY}:${PROJECT_NAME}-latest"
-    echo ""
-    echo "Please build and push your image first:"
-    echo "  ./do/submit"
-    echo ""
-    echo "After the build completes successfully, run this deploy script again."
-    exit 4
-fi
-
-echo "✅ ECR image found: ${ECR_REPOSITORY}:${PROJECT_NAME}-latest"
-IMAGE_TAG="${PROJECT_NAME}-latest"
-
-# ============================================================
-# Shared: Resolve secrets for container environment
-# ============================================================
-CONTAINER_ENV_JSON=""
-
-if [ -n "${HF_TOKEN_ARN:-}" ]; then
-    echo "🔐 Resolving HuggingFace token from Secrets Manager..."
-    RESOLVED_HF_TOKEN=$(aws secretsmanager get-secret-value --secret-id "${HF_TOKEN_ARN}" --query SecretString --output text --region "${AWS_REGION}") || {
-        echo "❌ Failed to resolve HuggingFace token from Secrets Manager"
-        exit 3
-    }
-    CONTAINER_ENV_JSON="\"HF_TOKEN\":\"${RESOLVED_HF_TOKEN}\""
-elif [ -n "${HF_TOKEN:-}" ]; then
-    CONTAINER_ENV_JSON="\"HF_TOKEN\":\"${HF_TOKEN}\""
-fi
-
-if [ -n "${NGC_API_KEY_ARN:-}" ]; then
-    echo "🔐 Resolving NGC API key from Secrets Manager..."
-    RESOLVED_NGC_KEY=$(aws secretsmanager get-secret-value --secret-id "${NGC_API_KEY_ARN}" --query SecretString --output text --region "${AWS_REGION}") || {
-        echo "❌ Failed to resolve NGC API key from Secrets Manager"
-        exit 3
-    }
-    if [ -n "${CONTAINER_ENV_JSON}" ]; then
-        CONTAINER_ENV_JSON="${CONTAINER_ENV_JSON},\"NGC_API_KEY\":\"${RESOLVED_NGC_KEY}\""
-    else
-        CONTAINER_ENV_JSON="\"NGC_API_KEY\":\"${RESOLVED_NGC_KEY}\""
-    fi
-elif [ -n "${NGC_API_KEY:-}" ]; then
-    if [ -n "${CONTAINER_ENV_JSON}" ]; then
-        CONTAINER_ENV_JSON="${CONTAINER_ENV_JSON},\"NGC_API_KEY\":\"${NGC_API_KEY}\""
-    else
-        CONTAINER_ENV_JSON="\"NGC_API_KEY\":\"${NGC_API_KEY}\""
-    fi
-fi
-
-<% if (deploymentTarget === 'realtime-inference') { %>
-# ============================================================
-# SageMaker Real-Time Inference Deployment (Inference Components)
-# ============================================================
-
-# Source shared helpers
-source "${SCRIPT_DIR}/lib/secrets.sh"
-source "${SCRIPT_DIR}/lib/wait.sh"
-source "${SCRIPT_DIR}/lib/endpoint-config.sh"
-source "${SCRIPT_DIR}/lib/inference-component.sh"
-
-# Validate execution role ARN
-if [ -z "${ROLE_ARN:-}" ]; then
-    echo "❌ Execution role ARN not provided"
-    echo ""
-    echo "Usage:"
-    echo "  export ROLE_ARN=arn:aws:iam::ACCOUNT_ID:role/YOUR_ROLE"
-    echo "  ./do/deploy"
-    echo ""
-    echo "Or set ROLE_ARN in do/config"
-    echo ""
-    echo "The execution role must have permissions for:"
-    echo "  • SageMaker endpoint and inference component management"
-    echo "  • ECR image access"
-    echo "  • S3 access (if using model artifacts)"
-    echo "  • CloudWatch Logs"
-    exit 3
-fi
-
-echo "   Using execution role: ${ROLE_ARN}"
-
-# Validate --ic argument if specified (set by --ic <name> or --force-ic <name>)
-if [ -n "${IC_TARGET}" ]; then
-    if [ ! -d "${SCRIPT_DIR}/ic" ]; then
-        echo "❌ IC name specified but no do/ic/ directory found"
-        echo "   This project does not use multi-IC configuration."
-        echo "   Remove --ic/--force-ic <name> to deploy using the legacy single-IC path."
-        exit 1
-    fi
-    if [ ! -f "${SCRIPT_DIR}/ic/${IC_TARGET}.conf" ]; then
-        echo "❌ IC config not found: do/ic/${IC_TARGET}.conf"
-        echo ""
-        echo "   Available ICs:"
-        for conf in "${SCRIPT_DIR}"/ic/*.conf; do
-            [ -f "${conf}" ] || continue
-            echo "     • $(basename "${conf}" .conf)"
-        done
-        echo ""
-        echo "   Usage: ./do/deploy --ic <name>"
-        exit 1
-    fi
-fi
-
-# Resolve container secrets (HF_TOKEN, NGC_API_KEY)
-resolve_secrets
-
-# ============================================================
-# Idempotency: check for existing deployment from a previous run
-# ============================================================
-SKIP_TO=""
-
-if [ "${FORCE_NEW}" = true ]; then
-    echo "🔄 --force: ignoring previous deployment, creating new resources."
-elif [ "${FORCE_IC}" = true ] && [ -n "${ENDPOINT_NAME:-}" ]; then
-    EP_STATUS=$(_get_endpoint_status "${ENDPOINT_NAME}")
-    if [ "${EP_STATUS}" = "InService" ]; then
-        if [ -n "${IC_TARGET}" ]; then
-            echo "🔄 --force-ic: recreating IC '${IC_TARGET}' on existing endpoint: ${ENDPOINT_NAME}"
-        else
-            echo "🔄 --force-ic: recreating ALL inference components on existing endpoint: ${ENDPOINT_NAME}"
-        fi
-        SKIP_TO="create_ic"
-    else
-        echo "⚠️  --force-ic requires an InService endpoint, but ${ENDPOINT_NAME} is: ${EP_STATUS:-not found}"
-        echo "   Use --force to create a new endpoint, or wait for the current one."
-        exit 4
-    fi
-elif [ -n "${ENDPOINT_NAME:-}" ]; then
-    echo "🔍 Checking for existing deployment: ${ENDPOINT_NAME}"
-
-    EP_STATUS=$(_get_endpoint_status "${ENDPOINT_NAME}")
-
-    case "${EP_STATUS}" in
-        InService)
-            echo "✅ Endpoint already InService: ${ENDPOINT_NAME}"
-
-            # Check inference component
-            if [ -n "${INFERENCE_COMPONENT_NAME:-}" ]; then
-                IC_STATUS=$(_get_ic_status "${INFERENCE_COMPONENT_NAME}")
-
-                case "${IC_STATUS}" in
-                    InService)
-                        echo "✅ Inference component already InService: ${INFERENCE_COMPONENT_NAME}"
-                        echo ""
-                        echo "📋 Deployment is already live. Nothing to do."
-                        echo "   Endpoint: ${ENDPOINT_NAME}"
-                        echo "   Inference Component: ${INFERENCE_COMPONENT_NAME}"
-                        echo ""
-                        echo "🧪 Test your endpoint:"
-                        echo "   ./do/test"
-                        echo ""
-                        echo "🧹 Clean up when done:"
-                        echo "   ./do/clean endpoint"
-                        exit 0
-                        ;;
-                    Creating)
-                        echo "⏳ Inference component still creating: ${INFERENCE_COMPONENT_NAME}"
-                        SKIP_TO="wait_ic"
-                        IC_DEPLOYED_NAME="${INFERENCE_COMPONENT_NAME}"
-                        ;;
-                    Failed)
-                        echo "⚠️  Inference component failed: ${INFERENCE_COMPONENT_NAME}"
-                        echo "   Will create a new inference component on the existing endpoint."
-                        SKIP_TO="create_ic"
-                        ;;
-                    *)
-                        # Stored IC not found — check if a different IC is running on this endpoint
-                        if [ "${ENDPOINT_EXTERNAL:-false}" = "true" ]; then
-                            # External endpoint: never adopt ICs we didn't create
-                            echo "   Stored IC not found on external endpoint. Will create a new one."
-                            SKIP_TO="create_ic"
-                        else
-                            LIVE_IC=$(_find_active_ic_on_endpoint "${ENDPOINT_NAME}")
-                            if [ -n "${LIVE_IC}" ] && [ "${LIVE_IC}" != "None" ]; then
-                                echo "✅ Found running inference component on endpoint: ${LIVE_IC}"
-                                echo "   (config had stale reference: ${INFERENCE_COMPONENT_NAME})"
-                                _update_config_var "INFERENCE_COMPONENT_NAME" "${LIVE_IC}"
-                                echo ""
-                                echo "📋 Deployment is already live. Nothing to do."
-                                echo "   Endpoint: ${ENDPOINT_NAME}"
-                                echo "   Inference Component: ${LIVE_IC}"
-                                echo ""
-                                echo "🧪 Test your endpoint:"
-                                echo "   ./do/test"
-                                echo ""
-                                echo "🧹 Clean up when done:"
-                                echo "   ./do/clean endpoint"
-                                exit 0
-                            else
-                                echo "   No existing inference component found on endpoint. Will create one."
-                                SKIP_TO="create_ic"
-                            fi
-                        fi
-                        ;;
-                esac
-            else
-                # No IC name in config — check if one is already running on the endpoint
-                if [ "${ENDPOINT_EXTERNAL:-false}" = "true" ]; then
-                    # External endpoint: never adopt ICs we didn't create
-                    echo "   No previous IC deployed by this project. Will create a new one."
-                    SKIP_TO="create_ic"
-                else
-                    LIVE_IC=$(_find_active_ic_on_endpoint "${ENDPOINT_NAME}")
-                    if [ -n "${LIVE_IC}" ] && [ "${LIVE_IC}" != "None" ]; then
-                        echo "✅ Found running inference component on endpoint: ${LIVE_IC}"
-                        _update_config_var "INFERENCE_COMPONENT_NAME" "${LIVE_IC}"
-                        echo ""
-                        echo "📋 Deployment is already live. Nothing to do."
-                        echo "   Endpoint: ${ENDPOINT_NAME}"
-                        echo "   Inference Component: ${LIVE_IC}"
-                        echo ""
-                        echo "🧪 Test your endpoint:"
-                        echo "   ./do/test"
-                        echo ""
-                        echo "🧹 Clean up when done:"
-                        echo "   ./do/clean endpoint"
-                        exit 0
-                    else
-                        SKIP_TO="create_ic"
-                    fi
-                fi
-            fi
-            ;;
-        Creating|Updating)
-            echo "⏳ Endpoint still ${EP_STATUS}: ${ENDPOINT_NAME}"
-            SKIP_TO="wait_endpoint"
-            ;;
-        Failed)
-            echo "⚠️  Previous endpoint failed: ${ENDPOINT_NAME}"
-            echo "   Creating a new deployment. Clean up the failed endpoint with:"
-            echo "   ./do/clean endpoint"
-            echo ""
-            # Fall through to create new resources
-            ;;
-        "")
-            echo "   Previous endpoint not found (may have been cleaned up). Creating new deployment."
-            ;;
-        *)
-            echo "   Endpoint in unexpected state: ${EP_STATUS}. Creating new deployment."
-            ;;
-    esac
-fi
-
-# ============================================================
-# Step 1: Create endpoint configuration and endpoint (skip if resuming)
-# ============================================================
-if [ -z "${SKIP_TO}" ]; then
-    if [ "${ENDPOINT_EXTERNAL:-false}" = "true" ]; then
-        # External endpoint: validate it still exists and is InService
-        echo "🔗 Using external endpoint: ${ENDPOINT_NAME}"
-        echo "   Validating endpoint status..."
-
-        EP_STATUS=$(_get_endpoint_status "${ENDPOINT_NAME}")
-
-        if [ -z "${EP_STATUS}" ]; then
-            echo "❌ External endpoint not found: ${ENDPOINT_NAME}"
-            echo "   The endpoint may have been deleted. Update ENDPOINT_NAME in do/config"
-            echo "   or remove ENDPOINT_EXTERNAL=true to create a new endpoint."
-            exit 4
-        fi
-
-        if [ "${EP_STATUS}" != "InService" ]; then
-            echo "❌ External endpoint not InService: ${ENDPOINT_NAME} (status: ${EP_STATUS})"
-            echo "   The endpoint must be InService before attaching inference components."
-            echo "   Wait for the endpoint to become InService, or update do/config."
-            exit 4
-        fi
-
-        echo "✅ External endpoint is InService: ${ENDPOINT_NAME}"
-        # Skip directly to IC creation — no endpoint config, no endpoint creation, no wait
-        SKIP_TO="create_ic"
-    else
-        TIMESTAMP=$(date +%s)
-        ENDPOINT_NAME="${PROJECT_NAME}-endpoint-${TIMESTAMP}"
-
-        _update_config_var "ENDPOINT_NAME" "${ENDPOINT_NAME}"
-
-        # Create endpoint configuration via shared helper
-        create_endpoint_config
-
-        _update_config_var "ENDPOINT_CONFIG_NAME" "${ENDPOINT_CONFIG_NAME}"
-
-        # Record endpoint config in manifest (non-blocking)
-        ENDPOINT_CONFIG_ARN="arn:aws:sagemaker:${AWS_REGION}:${AWS_ACCOUNT_ID}:endpoint-config/${ENDPOINT_CONFIG_NAME}"
-        ./do/manifest add \
-            --type sagemaker-endpoint-config \
-            --id "${ENDPOINT_CONFIG_ARN}" \
-            --project "${PROJECT_NAME}" \
-            --meta "{\"endpointConfigName\":\"${ENDPOINT_CONFIG_NAME}\",\"instanceType\":\"${INSTANCE_TYPE}\",\"region\":\"${AWS_REGION}\"}" \
-            2>/dev/null || true
-
-        # Step 2: Create endpoint
-        echo "🚀 Creating endpoint: ${ENDPOINT_NAME}"
-        if ! aws sagemaker create-endpoint \
-            --endpoint-name "${ENDPOINT_NAME}" \
-            --endpoint-config-name "${ENDPOINT_CONFIG_NAME}" \
-            --region "${AWS_REGION}"; then
-
-            echo "❌ Failed to create endpoint"
-            echo "   Check that:"
-            echo "   • Your IAM credentials have sagemaker:CreateEndpoint permission"
-            echo "   • You have sufficient service quota in region: ${AWS_REGION}"
-            exit 4
-        fi
-
-        echo "✅ Endpoint creation initiated: ${ENDPOINT_NAME}"
-
-        # Record endpoint in manifest (non-blocking)
-        ENDPOINT_ARN="arn:aws:sagemaker:${AWS_REGION}:${AWS_ACCOUNT_ID}:endpoint/${ENDPOINT_NAME}"
-        ./do/manifest add \
-            --type sagemaker-endpoint \
-            --id "${ENDPOINT_ARN}" \
-            --project "${PROJECT_NAME}" \
-            --meta "{\"endpointName\":\"${ENDPOINT_NAME}\",\"instanceType\":\"${INSTANCE_TYPE}\",\"region\":\"${AWS_REGION}\"}" \
-            2>/dev/null || true
-    fi
-fi
-
-# ============================================================
-# Wait for endpoint (skip if already InService or external)
-# ============================================================
-if [ -z "${SKIP_TO}" ] || [ "${SKIP_TO}" = "wait_endpoint" ]; then
-    echo "⏳ Waiting for endpoint to reach InService status..."
-    echo "   This may take a few minutes..."
-    echo "   If this times out, re-run ./do/deploy to resume."
-
-    wait_endpoint "${ENDPOINT_NAME}"
-
-    echo "✅ Endpoint is InService: ${ENDPOINT_NAME}"
-fi
-
-# ============================================================
-# Step 3: Deploy inference components (skip if resuming from wait_ic)
-# ============================================================
-if [ -z "${SKIP_TO}" ] || [ "${SKIP_TO}" = "create_ic" ] || [ "${SKIP_TO}" = "wait_endpoint" ]; then
-
-    if [ -d "${SCRIPT_DIR}/ic" ]; then
-        # _check_gpu_capacity
-        #   Best-effort capacity guardrail: sums IC_GPU_COUNT across all do/ic/*.conf
-        #   and compares against known GPU count for the instance type.
-        #   Warns (does not error) if total exceeds instance capacity.
-        #   Skips check if instance type is not in the known map.
-        _check_gpu_capacity() {
-            # Skip check if no INSTANCE_TYPE (external endpoints)
-            if [ -z "${INSTANCE_TYPE:-}" ]; then
-                return 0
-            fi
-
-            # Best-effort capacity guardrail: sums GPU requirements from base ICs only.
-            # NOTE: Only do/ic/*.conf files are counted. Adapter ICs (do/adapters/*.conf)
-            # share the base IC's GPU resources and have no ComputeResourceRequirements,
-            # so they are intentionally excluded from this capacity check.
-            #
-            # Hardcoded GPU counts for common SageMaker GPU instance types
-            local instance_gpus=""
-            case "${INSTANCE_TYPE}" in
-                ml.g4dn.xlarge)     instance_gpus=1 ;;
-                ml.g4dn.12xlarge)   instance_gpus=4 ;;
-                ml.g5.xlarge)       instance_gpus=1 ;;
-                ml.g5.2xlarge)      instance_gpus=1 ;;
-                ml.g5.4xlarge)      instance_gpus=1 ;;
-                ml.g5.8xlarge)      instance_gpus=1 ;;
-                ml.g5.12xlarge)     instance_gpus=4 ;;
-                ml.g5.48xlarge)     instance_gpus=8 ;;
-                ml.g6.xlarge)       instance_gpus=1 ;;
-                ml.g6.12xlarge)     instance_gpus=4 ;;
-                ml.g6.48xlarge)     instance_gpus=8 ;;
-                ml.g6e.xlarge)      instance_gpus=1 ;;
-                ml.g6e.2xlarge)     instance_gpus=1 ;;
-                ml.g6e.4xlarge)     instance_gpus=1 ;;
-                ml.g6e.8xlarge)     instance_gpus=1 ;;
-                ml.g6e.12xlarge)    instance_gpus=4 ;;
-                ml.g6e.48xlarge)    instance_gpus=8 ;;
-                ml.g7e.xlarge)      instance_gpus=1 ;;
-                ml.g7e.2xlarge)     instance_gpus=1 ;;
-                ml.g7e.4xlarge)     instance_gpus=1 ;;
-                ml.g7e.8xlarge)     instance_gpus=1 ;;
-                ml.g7e.12xlarge)    instance_gpus=4 ;;
-                ml.g7e.48xlarge)    instance_gpus=8 ;;
-                ml.p3.2xlarge)      instance_gpus=1 ;;
-                ml.p3.8xlarge)      instance_gpus=4 ;;
-                ml.p3.16xlarge)     instance_gpus=8 ;;
-                ml.p4d.24xlarge)    instance_gpus=8 ;;
-                ml.p4de.24xlarge)   instance_gpus=8 ;;
-                ml.p5.48xlarge)     instance_gpus=8 ;;
-                *)                  instance_gpus="" ;;
-            esac
-
-            # Skip check if instance type not in map
-            if [ -z "${instance_gpus}" ]; then
-                return 0
-            fi
-
-            # Sum IC_GPU_COUNT across all IC config files
-            local total_gpu_requested=0
-            for conf in "${SCRIPT_DIR}"/ic/*.conf; do
-                [ -f "${conf}" ] || continue
-                local ic_gpus
-                ic_gpus=$(grep "^export IC_GPU_COUNT=" "${conf}" 2>/dev/null | sed 's/^export IC_GPU_COUNT=//' | tr -d '"' || echo "1")
-                if [ -z "${ic_gpus}" ]; then
-                    ic_gpus=1
-                fi
-                total_gpu_requested=$(( total_gpu_requested + ic_gpus ))
-            done
-
-            if [ "${total_gpu_requested}" -gt "${instance_gpus}" ]; then
-                echo ""
-                echo "⚠️  GPU capacity warning: ICs request ${total_gpu_requested} GPUs total, but ${INSTANCE_TYPE} has ${instance_gpus} GPUs."
-                echo "   SageMaker will likely reject IC creation if capacity is exceeded."
-                echo "   Consider reducing IC_GPU_COUNT values or using a larger instance type."
-                echo ""
-            fi
-        }
-
-        # Run capacity guardrail before deploying ICs
-        _check_gpu_capacity
-
-        # _delete_and_wait_ic <ic_name>
-        #   Deletes an inference component and waits for deletion to complete.
-        #   Polls until the IC is no longer found (avoids name conflicts on recreate).
-        _delete_and_wait_ic() {
-            local ic_name="$1"
-            local delete_timeout=600  # 10 minutes max wait for deletion
-
-            echo "🗑️  Deleting inference component: ${ic_name}"
-            if ! aws sagemaker delete-inference-component \
-                --inference-component-name "${ic_name}" \
-                --region "${AWS_REGION}" 2>/dev/null; then
-                echo "   ⚠️  Delete call failed (IC may already be gone). Continuing..."
-                return 0
-            fi
-
-            echo "   Waiting for deletion to complete..."
-            local delete_start
-            delete_start=$(date +%s)
-
-            while true; do
-                local ic_status
-                ic_status=$(_get_ic_status "${ic_name}")
-
-                if [ -z "${ic_status}" ]; then
-                    echo "   ✅ Inference component deleted: ${ic_name}"
-                    break
-                fi
-
-                local elapsed=$(( $(date +%s) - delete_start ))
-                if [ "${elapsed}" -ge "${delete_timeout}" ]; then
-                    echo "   ⚠️  Deletion timed out after ${delete_timeout}s. IC status: ${ic_status}"
-                    echo "   Proceeding anyway — SageMaker may reject the new IC if name conflicts."
-                    break
-                fi
-
-                echo "   $(date +%H:%M:%S) Deleting... (${ic_status}, ${elapsed}s elapsed)"
-                sleep 15
-            done
-        }
-
-        # _deploy_single_ic <conf_file>
-        #   Deploys a single IC with per-IC idempotency:
-        #   - If FORCE_IC is true: delete existing IC, clear state, create fresh
-        #   - If IC_DEPLOYED_NAME is set and InService → skip
-        #   - If IC_DEPLOYED_NAME is set and Creating → wait for it
-        #   - If IC_DEPLOYED_NAME is set and Failed → recreate with new timestamp
-        #   - If IC_DEPLOYED_NAME is not set → create new IC
-        #   Fail-fast: exits immediately on failure.
-        _deploy_single_ic() {
-            local ic_conf="$1"
-            local ic_basename
-            ic_basename=$(basename "${ic_conf}" .conf)
-
-            # Source the IC config to check IC_DEPLOYED_NAME
-            # Use a subshell-safe approach: read the variable without polluting scope
-            local existing_ic_name=""
-            if grep -q "^export IC_DEPLOYED_NAME=" "${ic_conf}" 2>/dev/null; then
-                existing_ic_name=$(grep "^export IC_DEPLOYED_NAME=" "${ic_conf}" | sed 's/^export IC_DEPLOYED_NAME="//' | sed 's/"$//')
-            fi
-
-            # --force-ic: delete existing IC before recreating
-            if [ "${FORCE_IC}" = true ] && [ -n "${existing_ic_name}" ]; then
-                echo "🔄 --force-ic: recreating IC '${ic_basename}'"
-                _delete_and_wait_ic "${existing_ic_name}"
-
-                # Clear deployed state from config before recreating
-                _update_config_var "IC_DEPLOYED_NAME" "" "${ic_conf}"
-                _update_config_var "IC_DEPLOYED_AT" "" "${ic_conf}"
-                existing_ic_name=""
-            fi
-
-            if [ "${FORCE_IC}" = true ] && [ -z "${existing_ic_name}" ]; then
-                # Force mode with no existing IC — just create new
-                create_inference_component "${ic_conf}"
-            elif [ -n "${existing_ic_name}" ]; then
-                # IC was previously deployed — check its current status
-                local ic_status
-                ic_status=$(_get_ic_status "${existing_ic_name}")
-
-                case "${ic_status}" in
-                    InService)
-                        echo "✅ IC '${ic_basename}' already InService: ${existing_ic_name} — skipping"
-                        IC_DEPLOYED_NAME="${existing_ic_name}"
-                        return 0
-                        ;;
-                    Creating)
-                        echo "⏳ IC '${ic_basename}' is still Creating: ${existing_ic_name} — waiting..."
-                        IC_DEPLOYED_NAME="${existing_ic_name}"
-                        wait_ic "${IC_DEPLOYED_NAME}"
-                        echo "✅ Inference component is InService: ${IC_DEPLOYED_NAME}"
-                        return 0
-                        ;;
-                    Failed)
-                        echo "⚠️  IC '${ic_basename}' previously Failed: ${existing_ic_name} — recreating..."
-                        create_inference_component "${ic_conf}"
-                        ;;
-                    *)
-                        echo "   IC '${ic_basename}' has unknown/missing status for ${existing_ic_name} — creating new..."
-                        create_inference_component "${ic_conf}"
-                        ;;
-                esac
-            else
-                # No previous deployment — create new IC
-                create_inference_component "${ic_conf}"
-            fi
-
-            echo "⏳ Waiting for inference component to reach InService status..."
-            echo "   This may take 5-10 minutes..."
-
-            wait_ic "${IC_DEPLOYED_NAME}"
-
-            echo "✅ Inference component is InService: ${IC_DEPLOYED_NAME}"
-
-            # Record inference component in manifest (non-blocking)
-            local ic_arn="arn:aws:sagemaker:${AWS_REGION}:${AWS_ACCOUNT_ID}:inference-component/${IC_DEPLOYED_NAME}"
-            ./do/manifest add \
-                --type sagemaker-inference-component \
-                --id "${ic_arn}" \
-                --project "${PROJECT_NAME}" \
-                --meta "{\"inferenceComponentName\":\"${IC_DEPLOYED_NAME}\",\"endpointName\":\"${ENDPOINT_NAME}\",\"instanceType\":\"${INSTANCE_TYPE:-external}\",\"region\":\"${AWS_REGION}\"}" \
-                2>/dev/null || true
-        }
-
-        if [ -n "${IC_TARGET}" ]; then
-            # Single IC path: deploy only the named IC
-            echo ""
-            echo "── Deploying IC: ${IC_TARGET} ──"
-            _deploy_single_ic "${SCRIPT_DIR}/ic/${IC_TARGET}.conf"
-        else
-            # Multi-IC path: iterate all IC config files (alphabetical order)
-            IC_SUMMARY=""
-            IC_DEPLOY_FAILED=false
-
-            for conf in "${SCRIPT_DIR}"/ic/*.conf; do
-                [ -f "${conf}" ] || continue
-                local_ic_basename=$(basename "${conf}" .conf)
-                echo ""
-                echo "── Deploying IC: ${local_ic_basename} ──"
-
-                if ! _deploy_single_ic "${conf}"; then
-                    echo "❌ IC '${local_ic_basename}' failed to deploy. Stopping."
-                    IC_SUMMARY="${IC_SUMMARY}   ${local_ic_basename}: FAILED\n"
-                    IC_DEPLOY_FAILED=true
-                    break
-                fi
-
-                IC_SUMMARY="${IC_SUMMARY}   ${local_ic_basename}: ${IC_DEPLOYED_NAME} [InService]\n"
-            done
-
-            # Print summary
-            echo ""
-            echo "📋 IC Deployment Summary:"
-            echo -e "${IC_SUMMARY}"
-
-            if [ "${IC_DEPLOY_FAILED}" = true ]; then
-                echo "❌ Deployment stopped due to IC failure. Fix the issue and re-run ./do/deploy to resume."
-                exit 4
-            fi
-        fi
-    else
-        # Legacy single-IC path: no do/ic/ directory
-        create_inference_component_legacy
-
-        echo "⏳ Waiting for inference component to reach InService status..."
-        echo "   This may take 5-10 minutes..."
-        echo "   If this times out, re-run ./do/deploy to resume."
-
-        wait_ic "${IC_DEPLOYED_NAME}"
-
-        echo "✅ Inference component is InService: ${IC_DEPLOYED_NAME}"
-
-        # Record inference component in manifest (non-blocking)
-        IC_ARN="arn:aws:sagemaker:${AWS_REGION}:${AWS_ACCOUNT_ID}:inference-component/${IC_DEPLOYED_NAME}"
-        ./do/manifest add \
-            --type sagemaker-inference-component \
-            --id "${IC_ARN}" \
-            --project "${PROJECT_NAME}" \
-            --meta "{\"inferenceComponentName\":\"${IC_DEPLOYED_NAME}\",\"endpointName\":\"${ENDPOINT_NAME}\",\"instanceType\":\"${INSTANCE_TYPE:-external}\",\"region\":\"${AWS_REGION}\"}" \
-            2>/dev/null || true
-    fi
-
-elif [ "${SKIP_TO}" = "wait_ic" ]; then
-    # Resuming: just wait for the IC that was already being created
-    echo "⏳ Waiting for inference component to reach InService status..."
-    echo "   This may take 5-10 minutes..."
-    echo "   If this times out, re-run ./do/deploy to resume."
-
-    wait_ic "${IC_DEPLOYED_NAME}"
-
-    echo "✅ Inference component is InService: ${IC_DEPLOYED_NAME}"
-fi
-
-echo "✅ Deployment complete!"
-echo ""
-echo "📋 Deployment Details:"
-echo "   Endpoint: ${ENDPOINT_NAME}"
-if [ "${ENDPOINT_EXTERNAL:-false}" = "true" ]; then
-    echo "   Endpoint Config: (external — not managed by this project)"
-    echo "   Region: ${AWS_REGION}"
-else
-    echo "   Endpoint Config: ${ENDPOINT_CONFIG_NAME:-N/A}"
-    echo "   Region: ${AWS_REGION}"
-    echo "   Instance Type: ${INSTANCE_TYPE}"
-fi
-echo "   Image: ${ECR_REPOSITORY}:${IMAGE_TAG}"
-echo ""
-echo "📋 What's next?"
-echo "   • Test your endpoint:         ./do/test"
-<% if (typeof includeBenchmark !== 'undefined' && includeBenchmark) { %>
-echo "   • Benchmark performance:      ./do/benchmark"
-<% } %>
-<% if (typeof enableLora !== 'undefined' && enableLora) { %>
-echo "   • Add a LoRA adapter:         ./do/adapter add <name> --weights s3://..."
-<% } %>
-echo "   • View endpoint status:       ./do/status"
-echo "   • Register this deployment:   ./do/register"
-echo "   • View logs:                  ./do/logs"
-<% if (!(typeof existingEndpointName !== 'undefined' && existingEndpointName)) { %>
-echo "   • Clean up when done:         ./do/clean endpoint"
-<% } %>
-
-<% } else if (deploymentTarget === 'async-inference') { %>
-# ============================================================
-# SageMaker Async Inference Deployment (Model-Based)
-# SageMaker async inference does NOT support Inference Components.
-# Flow: create-model → create-endpoint-config (with AsyncInferenceConfig) → create-endpoint
-# ============================================================
-
-# Source shared helpers
-source "${SCRIPT_DIR}/lib/secrets.sh"
-source "${SCRIPT_DIR}/lib/wait.sh"
-
-# Resolve container secrets (HF_TOKEN, NGC_API_KEY)
-resolve_secrets
-
-# Validate execution role ARN
-if [ -z "${ROLE_ARN:-}" ]; then
-    echo "❌ Execution role ARN not provided"
-    echo ""
-    echo "Usage:"
-    echo "  export ROLE_ARN=arn:aws:iam::ACCOUNT_ID:role/YOUR_ROLE"
-    echo "  ./do/deploy"
-    echo ""
-    echo "Or set ROLE_ARN in do/config"
-    echo ""
-    echo "The execution role must have permissions for:"
-    echo "  • SageMaker model and endpoint management"
-    echo "  • ECR image access"
-    echo "  • S3 write access for async output path: ${ASYNC_S3_OUTPUT_PATH}"
-    echo "  • SNS publish permissions (optional, for notifications)"
-    echo "  • CloudWatch Logs"
-    exit 3
-fi
-
-echo "   Using execution role: ${ROLE_ARN}"
-
-# ============================================================
-# Bootstrap async infrastructure (S3 bucket + SNS topics)
-# ============================================================
-
-# Extract bucket name from S3 output path
-ASYNC_S3_BUCKET=$(echo "${ASYNC_S3_OUTPUT_PATH}" | sed 's|s3://||' | cut -d'/' -f1)
-
-<% if (!asyncS3OutputPath) { %>
-# Bootstrap default S3 bucket (check-and-create)
-echo "🔍 Checking if S3 bucket exists: ${ASYNC_S3_BUCKET}"
-if ! aws s3api head-bucket --bucket "${ASYNC_S3_BUCKET}" --region "${AWS_REGION}" 2>/dev/null; then
-    echo "📦 Creating S3 bucket: ${ASYNC_S3_BUCKET}"
-    if [ "${AWS_REGION}" = "us-east-1" ]; then
-        if ! aws s3api create-bucket \
-            --bucket "${ASYNC_S3_BUCKET}" \
-            --region "${AWS_REGION}"; then
-            echo "❌ Failed to create S3 bucket: ${ASYNC_S3_BUCKET}"
-            echo ""
-            echo "   Check that:"
-            echo "   • Your IAM credentials have s3:CreateBucket permission"
-            echo "   • The bucket name is not already taken globally"
-            exit 4
-        fi
-    else
-        if ! aws s3api create-bucket \
-            --bucket "${ASYNC_S3_BUCKET}" \
-            --region "${AWS_REGION}" \
-            --create-bucket-configuration LocationConstraint="${AWS_REGION}"; then
-            echo "❌ Failed to create S3 bucket: ${ASYNC_S3_BUCKET}"
-            echo ""
-            echo "   Check that:"
-            echo "   • Your IAM credentials have s3:CreateBucket permission"
-            echo "   • The bucket name is not already taken globally"
-            exit 4
-        fi
-    fi
-    echo "✅ S3 bucket created: ${ASYNC_S3_BUCKET}"
-else
-    echo "✅ S3 bucket exists: ${ASYNC_S3_BUCKET}"
-fi
-<% } else { %>
-# Custom S3 output path provided — skip bucket creation
-echo "✅ Using custom S3 output path: ${ASYNC_S3_OUTPUT_PATH}"
-<% } %>
-
-# Extract topic name from SNS success topic ARN
-ASYNC_SNS_SUCCESS_TOPIC_NAME=$(echo "${ASYNC_SNS_SUCCESS_TOPIC}" | awk -F: '{print $NF}')
-
-<% if (!asyncSnsSuccessTopic) { %>
-# Bootstrap default SNS success topic (check-and-create)
-echo "🔍 Checking if SNS success topic exists: ${ASYNC_SNS_SUCCESS_TOPIC_NAME}"
-if ! aws sns get-topic-attributes --topic-arn "${ASYNC_SNS_SUCCESS_TOPIC}" --region "${AWS_REGION}" 2>/dev/null; then
-    echo "📦 Creating SNS success topic: ${ASYNC_SNS_SUCCESS_TOPIC_NAME}"
-    if ! aws sns create-topic \
-        --name "${ASYNC_SNS_SUCCESS_TOPIC_NAME}" \
-        --region "${AWS_REGION}" > /dev/null; then
-        echo "❌ Failed to create SNS success topic: ${ASYNC_SNS_SUCCESS_TOPIC_NAME}"
-        echo ""
-        echo "   Check that:"
-        echo "   • Your IAM credentials have sns:CreateTopic permission"
-        exit 4
-    fi
-    echo "✅ SNS success topic created: ${ASYNC_SNS_SUCCESS_TOPIC_NAME}"
-else
-    echo "✅ SNS success topic exists: ${ASYNC_SNS_SUCCESS_TOPIC_NAME}"
-fi
-
-# Record SNS success topic in manifest (non-blocking)
-./do/manifest add \
-    --type sns-topic \
-    --id "${ASYNC_SNS_SUCCESS_TOPIC}" \
-    --project "${PROJECT_NAME}" \
-    --meta "{\"topicName\":\"${ASYNC_SNS_SUCCESS_TOPIC_NAME}\",\"purpose\":\"async-success\",\"region\":\"${AWS_REGION}\"}" \
-    2>/dev/null || true
-
-<% } else { %>
-# Custom SNS success topic ARN provided — skip topic creation
-echo "✅ Using custom SNS success topic: ${ASYNC_SNS_SUCCESS_TOPIC}"
-
-# Record SNS success topic in manifest (non-blocking)
-ASYNC_SNS_SUCCESS_TOPIC_NAME=$(echo "${ASYNC_SNS_SUCCESS_TOPIC}" | awk -F: '{print $NF}')
-./do/manifest add \
-    --type sns-topic \
-    --id "${ASYNC_SNS_SUCCESS_TOPIC}" \
-    --project "${PROJECT_NAME}" \
-    --meta "{\"topicName\":\"${ASYNC_SNS_SUCCESS_TOPIC_NAME}\",\"purpose\":\"async-success\",\"region\":\"${AWS_REGION}\"}" \
-    2>/dev/null || true
-
-<% } %>
-
-# Extract topic name from SNS error topic ARN
-ASYNC_SNS_ERROR_TOPIC_NAME=$(echo "${ASYNC_SNS_ERROR_TOPIC}" | awk -F: '{print $NF}')
-
-<% if (!asyncSnsErrorTopic) { %>
-# Bootstrap default SNS error topic (check-and-create)
-echo "🔍 Checking if SNS error topic exists: ${ASYNC_SNS_ERROR_TOPIC_NAME}"
-if ! aws sns get-topic-attributes --topic-arn "${ASYNC_SNS_ERROR_TOPIC}" --region "${AWS_REGION}" 2>/dev/null; then
-    echo "📦 Creating SNS error topic: ${ASYNC_SNS_ERROR_TOPIC_NAME}"
-    if ! aws sns create-topic \
-        --name "${ASYNC_SNS_ERROR_TOPIC_NAME}" \
-        --region "${AWS_REGION}" > /dev/null; then
-        echo "❌ Failed to create SNS error topic: ${ASYNC_SNS_ERROR_TOPIC_NAME}"
-        echo ""
-        echo "   Check that:"
-        echo "   • Your IAM credentials have sns:CreateTopic permission"
-        exit 4
-    fi
-    echo "✅ SNS error topic created: ${ASYNC_SNS_ERROR_TOPIC_NAME}"
-else
-    echo "✅ SNS error topic exists: ${ASYNC_SNS_ERROR_TOPIC_NAME}"
-fi
-
-# Record SNS error topic in manifest (non-blocking)
-./do/manifest add \
-    --type sns-topic \
-    --id "${ASYNC_SNS_ERROR_TOPIC}" \
-    --project "${PROJECT_NAME}" \
-    --meta "{\"topicName\":\"${ASYNC_SNS_ERROR_TOPIC_NAME}\",\"purpose\":\"async-error\",\"region\":\"${AWS_REGION}\"}" \
-    2>/dev/null || true
-
-<% } else { %>
-# Custom SNS error topic ARN provided — skip topic creation
-echo "✅ Using custom SNS error topic: ${ASYNC_SNS_ERROR_TOPIC}"
-
-# Record SNS error topic in manifest (non-blocking)
-ASYNC_SNS_ERROR_TOPIC_NAME=$(echo "${ASYNC_SNS_ERROR_TOPIC}" | awk -F: '{print $NF}')
-./do/manifest add \
-    --type sns-topic \
-    --id "${ASYNC_SNS_ERROR_TOPIC}" \
-    --project "${PROJECT_NAME}" \
-    --meta "{\"topicName\":\"${ASYNC_SNS_ERROR_TOPIC_NAME}\",\"purpose\":\"async-error\",\"region\":\"${AWS_REGION}\"}" \
-    2>/dev/null || true
-
-<% } %>
-
-# ============================================================
-# Create async endpoint (classic model-based flow)
-# SageMaker async inference does NOT support Inference Components.
-# Flow: create-model → create-endpoint-config (with AsyncInferenceConfig) → create-endpoint
-# ============================================================
-
-# ============================================================
-# Idempotency: check for existing deployment from a previous run
-# ============================================================
-SKIP_TO=""
-
-if [ "${FORCE_NEW}" = true ]; then
-    echo "🔄 --force: ignoring previous deployment, creating new resources."
-elif [ -n "${ENDPOINT_NAME:-}" ]; then
-    echo "🔍 Checking for existing deployment: ${ENDPOINT_NAME}"
-
-    EP_STATUS=$(_get_endpoint_status "${ENDPOINT_NAME}")
-
-    case "${EP_STATUS}" in
-        InService)
-            echo "✅ Async endpoint already InService: ${ENDPOINT_NAME}"
-            echo ""
-            echo "📋 Deployment is already live. Nothing to do."
-            echo "   Endpoint: ${ENDPOINT_NAME}"
-            echo ""
-            echo "🧪 Test your async endpoint:"
-            echo "   ./do/test"
-            echo ""
-            echo "🧹 Clean up when done:"
-            echo "   ./do/clean endpoint"
-            exit 0
-            ;;
-        Creating|Updating)
-            echo "⏳ Endpoint still ${EP_STATUS}: ${ENDPOINT_NAME}"
-            SKIP_TO="wait_endpoint"
-            ;;
-        Failed)
-            echo "⚠️  Previous endpoint failed: ${ENDPOINT_NAME}"
-            echo "   Creating a new deployment. Clean up the failed endpoint with:"
-            echo "   ./do/clean endpoint"
-            echo ""
-            ;;
-        "")
-            echo "   Previous endpoint not found (may have been cleaned up). Creating new deployment."
-            ;;
-        *)
-            echo "   Endpoint in unexpected state: ${EP_STATUS}. Creating new deployment."
-            ;;
-    esac
-fi
-
-# ============================================================
-# Create async resources (skip if resuming from wait)
-# ============================================================
-if [ -z "${SKIP_TO}" ]; then
-    TIMESTAMP=$(date +%s)
-    MODEL_NAME_SM="${PROJECT_NAME}-async-model-${TIMESTAMP}"
-    ENDPOINT_CONFIG_NAME="${PROJECT_NAME}-async-epc-${TIMESTAMP}"
-    ENDPOINT_NAME="${PROJECT_NAME}-async-ep-${TIMESTAMP}"
-
-    _update_config_var "ENDPOINT_NAME" "${ENDPOINT_NAME}"
-    _update_config_var "ENDPOINT_CONFIG_NAME" "${ENDPOINT_CONFIG_NAME}"
-    _update_config_var "SAGEMAKER_MODEL_NAME" "${MODEL_NAME_SM}"
-
-    # Step 1: Create SageMaker model
-    # Build primary container spec
-    PRIMARY_CONTAINER="{\"Image\":\"${ECR_REPOSITORY}:${IMAGE_TAG}\""
-    if [ -n "${CONTAINER_ENV_JSON}" ]; then
-        PRIMARY_CONTAINER="${PRIMARY_CONTAINER},\"Environment\":{${CONTAINER_ENV_JSON}}"
-    fi
-    PRIMARY_CONTAINER="${PRIMARY_CONTAINER}}"
-
-    echo "📦 Creating SageMaker model: ${MODEL_NAME_SM}"
-    if ! aws sagemaker create-model \
-        --model-name "${MODEL_NAME_SM}" \
-        --primary-container "${PRIMARY_CONTAINER}" \
-        --execution-role-arn "${ROLE_ARN}" \
-        --region "${AWS_REGION}"; then
-
-        echo "❌ Failed to create SageMaker model"
-        echo "   Check that:"
-        echo "   • The execution role ARN is valid"
-        echo "   • The ECR image exists and is accessible"
-        echo "   • The IAM role has ecr:GetDownloadUrlForLayer permission"
-        exit 4
-    fi
-
-    echo "✅ SageMaker model created: ${MODEL_NAME_SM}"
-
-    # Record model in manifest (non-blocking)
-    MODEL_ARN="arn:aws:sagemaker:${AWS_REGION}:${AWS_ACCOUNT_ID}:model/${MODEL_NAME_SM}"
-    ./do/manifest add \
-        --type sagemaker-model \
-        --id "${MODEL_ARN}" \
-        --project "${PROJECT_NAME}" \
-        --meta "{\"modelName\":\"${MODEL_NAME_SM}\",\"region\":\"${AWS_REGION}\"}" \
-        2>/dev/null || true
-
-    # Build production variant JSON (classic: includes ModelName, no execution-role-arn on endpoint config)
-    VARIANT_JSON="[{\"VariantName\":\"AllTraffic\",\"ModelName\":\"${MODEL_NAME_SM}\",\"InstanceType\":\"${INSTANCE_TYPE}\",\"InitialInstanceCount\":1"
-
-    # Append InferenceAmiVersion if configured
-    if [ -n "${INFERENCE_AMI_VERSION:-}" ]; then
-        VARIANT_JSON="${VARIANT_JSON},\"InferenceAmiVersion\":\"${INFERENCE_AMI_VERSION}\""
-        echo "   AMI version: ${INFERENCE_AMI_VERSION}"
-    fi
-
-    VARIANT_JSON="${VARIANT_JSON}}]"
-
-    # Build AsyncInferenceConfig JSON
-    ASYNC_CONFIG="{\"OutputConfig\":{\"S3OutputPath\":\"${ASYNC_S3_OUTPUT_PATH}\",\"NotificationConfig\":{\"SuccessTopic\":\"${ASYNC_SNS_SUCCESS_TOPIC}\",\"ErrorTopic\":\"${ASYNC_SNS_ERROR_TOPIC}\"}}"
-
-    if [ -n "${ASYNC_MAX_CONCURRENT_INVOCATIONS:-}" ]; then
-        ASYNC_CONFIG="${ASYNC_CONFIG},\"ClientConfig\":{\"MaxConcurrentInvocationsPerInstance\":${ASYNC_MAX_CONCURRENT_INVOCATIONS}}"
-    fi
-
-    ASYNC_CONFIG="${ASYNC_CONFIG}}"
-
-    # Step 2: Create endpoint configuration with AsyncInferenceConfig (no --execution-role-arn)
-    echo "⚙️  Creating async endpoint configuration: ${ENDPOINT_CONFIG_NAME}"
-    if ! aws sagemaker create-endpoint-config \
-        --endpoint-config-name "${ENDPOINT_CONFIG_NAME}" \
-        --production-variants "${VARIANT_JSON}" \
-        --async-inference-config "${ASYNC_CONFIG}" \
-        --region "${AWS_REGION}"; then
-
-        echo "❌ Failed to create async endpoint configuration"
-        echo "   Check that:"
-        echo "   • The S3 output path is accessible: ${ASYNC_S3_OUTPUT_PATH}"
-        echo "   • The IAM role has s3:PutObject permission on the output path"
-        echo "   • The instance type is valid: ${INSTANCE_TYPE}"
-        echo "   • The instance type is available in region: ${AWS_REGION}"
-        echo "   • You have sufficient service quota for the instance type"
-        exit 4
-    fi
-
-    echo "✅ Async endpoint configuration created: ${ENDPOINT_CONFIG_NAME}"
-
-    # Record endpoint config in manifest (non-blocking)
-    ENDPOINT_CONFIG_ARN="arn:aws:sagemaker:${AWS_REGION}:${AWS_ACCOUNT_ID}:endpoint-config/${ENDPOINT_CONFIG_NAME}"
-    ./do/manifest add \
-        --type sagemaker-endpoint-config \
-        --id "${ENDPOINT_CONFIG_ARN}" \
-        --project "${PROJECT_NAME}" \
-        --meta "{\"endpointConfigName\":\"${ENDPOINT_CONFIG_NAME}\",\"instanceType\":\"${INSTANCE_TYPE}\",\"region\":\"${AWS_REGION}\"}" \
-        2>/dev/null || true
-
-    # Step 3: Create endpoint
-    echo "🚀 Creating async endpoint: ${ENDPOINT_NAME}"
-    if ! aws sagemaker create-endpoint \
-        --endpoint-name "${ENDPOINT_NAME}" \
-        --endpoint-config-name "${ENDPOINT_CONFIG_NAME}" \
-        --region "${AWS_REGION}"; then
-
-        echo "❌ Failed to create async endpoint"
-        echo "   Check that:"
-        echo "   • Your IAM credentials have sagemaker:CreateEndpoint permission"
-        echo "   • You have sufficient service quota in region: ${AWS_REGION}"
-        exit 4
-    fi
-
-    echo "✅ Async endpoint creation initiated: ${ENDPOINT_NAME}"
-
-    # Record endpoint in manifest (non-blocking)
-    ENDPOINT_ARN="arn:aws:sagemaker:${AWS_REGION}:${AWS_ACCOUNT_ID}:endpoint/${ENDPOINT_NAME}"
-    ./do/manifest add \
-        --type sagemaker-endpoint \
-        --id "${ENDPOINT_ARN}" \
-        --project "${PROJECT_NAME}" \
-        --meta "{\"endpointName\":\"${ENDPOINT_NAME}\",\"instanceType\":\"${INSTANCE_TYPE}\",\"region\":\"${AWS_REGION}\"}" \
-        2>/dev/null || true
-fi
-
-# ============================================================
-# Wait for endpoint (skip if already InService)
-# ============================================================
-if [ -z "${SKIP_TO}" ] || [ "${SKIP_TO}" = "wait_endpoint" ]; then
-    echo "⏳ Waiting for async endpoint to reach InService status..."
-    echo "   This may take several minutes..."
-    echo "   If this times out, re-run ./do/deploy to resume."
-
-    wait_endpoint "${ENDPOINT_NAME}"
-fi
-
-echo "✅ Async deployment complete!"
-echo ""
-echo "📋 Deployment Details:"
-echo "   Endpoint: ${ENDPOINT_NAME}"
-echo "   Endpoint Config: ${ENDPOINT_CONFIG_NAME}"
-echo "   Model: ${MODEL_NAME_SM}"
-echo "   Region: ${AWS_REGION}"
-echo "   Instance Type: ${INSTANCE_TYPE}"
-echo "   Image: ${ECR_REPOSITORY}:${IMAGE_TAG}"
-echo "   S3 Output: ${ASYNC_S3_OUTPUT_PATH}"
-echo "   SNS Success: ${ASYNC_SNS_SUCCESS_TOPIC}"
-echo "   SNS Error: ${ASYNC_SNS_ERROR_TOPIC}"
-echo ""
-echo "📋 What's next?"
-echo "   • Test your async endpoint:   ./do/test"
-echo "   • Check async output:         aws s3 ls ${ASYNC_S3_OUTPUT_PATH}"
-<% if (typeof includeBenchmark !== 'undefined' && includeBenchmark) { %>
-echo "   • Benchmark performance:      ./do/benchmark"
-<% } %>
-echo "   • Register this deployment:   ./do/register"
-echo "   • View logs:                  ./do/logs"
-echo "   • Clean up when done:         ./do/clean endpoint"
-
-<% } else if (deploymentTarget === 'hyperpod-eks') { %>
-# ============================================================
-# HyperPod EKS Deployment
-# ============================================================
-
-# Get kubeconfig for HyperPod cluster
-echo "🔑 Configuring kubectl for HyperPod cluster..."
-KUBECONFIG_PATH="${HOME}/.kube/hyperpod-${HYPERPOD_CLUSTER_NAME}"
-
-# Step 1: Describe the HyperPod cluster to get the underlying EKS cluster ARN
-EKS_CLUSTER_ARN=$(aws sagemaker describe-cluster \
-    --cluster-name "${HYPERPOD_CLUSTER_NAME}" \
-    --region "${AWS_REGION}" \
-    --query "Orchestrator.Eks.ClusterArn" \
-    --output text 2>&1) || {
-    echo "❌ Failed to describe HyperPod cluster: ${HYPERPOD_CLUSTER_NAME}"
-    echo ""
-    echo "   Error details:"
-    echo "   ${EKS_CLUSTER_ARN}"
-    echo ""
-    echo "   Check that:"
-    echo "   • The cluster name is correct"
-    echo "   • The cluster exists in region: ${AWS_REGION}"
-    echo "   • Your IAM user/role has permission to access the cluster"
-    echo ""
-    echo "   Required IAM permissions:"
-    echo "   • sagemaker:DescribeCluster"
-    echo "   • eks:DescribeCluster"
-    exit 4
-}
-
-# Step 2: Extract the EKS cluster name from the ARN
-EKS_CLUSTER_NAME=$(echo "${EKS_CLUSTER_ARN}" | awk -F'/' '{print $NF}')
-echo "   HyperPod cluster: ${HYPERPOD_CLUSTER_NAME}"
-echo "   EKS cluster: ${EKS_CLUSTER_NAME}"
-
-# Step 3: Update kubeconfig using the EKS cluster
-if ! aws eks update-kubeconfig \
-    --name "${EKS_CLUSTER_NAME}" \
-    --region "${AWS_REGION}" \
-    --kubeconfig "${KUBECONFIG_PATH}" 2>&1; then
-    echo "❌ Failed to configure kubectl for EKS cluster: ${EKS_CLUSTER_NAME}"
-    echo ""
-    echo "   Required IAM permissions:"
-    echo "   • eks:DescribeCluster"
-    echo "   • eks:AccessKubernetesApi"
-    exit 4
-fi
-
-export KUBECONFIG="${KUBECONFIG_PATH}"
-echo "✅ Kubeconfig saved to: ${KUBECONFIG_PATH}"
-
-# Verify cluster connectivity
-echo "🔍 Verifying cluster connectivity..."
-if ! kubectl cluster-info &> /dev/null; then
-    echo "❌ Cannot connect to HyperPod cluster"
-    echo ""
-    echo "   Check that:"
-    echo "   • The cluster is in 'InService' status"
-    echo "   • Your network can reach the cluster API server"
-    echo "   • Your IAM credentials are valid"
-    exit 4
-fi
-echo "✅ Connected to HyperPod cluster"
-
-# Create namespace if it doesn't exist
-echo "📁 Ensuring namespace exists: ${HYPERPOD_NAMESPACE}"
-if ! kubectl create namespace "${HYPERPOD_NAMESPACE}" --dry-run=client -o yaml | kubectl apply -f - 2>&1; then
-    echo "⚠️  Warning: Could not create/verify namespace"
-fi
-
-# Apply Kubernetes manifests
-echo "📄 Applying Kubernetes manifests from hyperpod/..."
-
-# Substitute shell variables (e.g. ${AWS_ACCOUNT_ID}) in manifests before applying
-export AWS_ACCOUNT_ID
-export ECR_IMAGE="${ECR_REPOSITORY}:${IMAGE_TAG}"
-
-APPLY_OUTPUT=""
-APPLY_EXIT_CODE=0
-for manifest in hyperpod/*.yaml; do
-    # Skip files that contain no Kubernetes objects (e.g. comment-only PVC stubs)
-    RENDERED=$(envsubst < "${manifest}")
-    if echo "${RENDERED}" | grep -q '^kind:'; then
-        FILE_OUTPUT=$(echo "${RENDERED}" | kubectl apply -n "${HYPERPOD_NAMESPACE}" -f - 2>&1) || {
-            APPLY_EXIT_CODE=$?
-        }
-        APPLY_OUTPUT="${APPLY_OUTPUT}${FILE_OUTPUT}\n"
-    fi
-done
-
-if [ "${APPLY_EXIT_CODE}" -ne 0 ]; then
-    echo ""
-    echo "❌ Failed to apply Kubernetes manifests"
-    echo ""
-    echo "   Error details:"
-    echo "   ${APPLY_OUTPUT}"
-    echo ""
-    echo "   Common issues:"
-    echo "   • Insufficient node capacity - check available GPU nodes"
-    echo "   • Resource requests exceed node capacity"
-    echo "   • RBAC permissions - ensure you have permission to create resources in namespace '${HYPERPOD_NAMESPACE}'"
-    echo "   • Invalid manifest syntax"
-<% if (fsxVolumeHandle) { %>
-    echo "   • PVC creation failure - verify the FSx CSI driver is installed on the cluster"
-    echo "     kubectl get csidriver -o name | grep fsx"
-<% } %>
-    echo ""
-    echo "   Debug commands:"
-    echo "   kubectl get nodes -o wide"
-    echo "   kubectl describe nodes"
-    echo "   kubectl get events -n ${HYPERPOD_NAMESPACE}"
-    exit ${APPLY_EXIT_CODE}
-fi
-
-echo "✅ Kubernetes manifests applied"
-
-# Record k8s deployment and service in manifest (non-blocking)
-./do/manifest add \
-    --type k8s-deployment \
-    --id "${HYPERPOD_NAMESPACE}/${PROJECT_NAME}" \
-    --project "${PROJECT_NAME}" \
-    --meta "{\"namespace\":\"${HYPERPOD_NAMESPACE}\",\"deploymentName\":\"${PROJECT_NAME}\",\"clusterName\":\"${HYPERPOD_CLUSTER_NAME}\",\"region\":\"${AWS_REGION}\"}" \
-    2>/dev/null || true
-
-./do/manifest add \
-    --type k8s-service \
-    --id "${HYPERPOD_NAMESPACE}/${PROJECT_NAME}" \
-    --project "${PROJECT_NAME}" \
-    --meta "{\"namespace\":\"${HYPERPOD_NAMESPACE}\",\"serviceName\":\"${PROJECT_NAME}\",\"clusterName\":\"${HYPERPOD_CLUSTER_NAME}\",\"region\":\"${AWS_REGION}\"}" \
-    2>/dev/null || true
-
-# Wait for deployment to be ready
-DEPLOY_TIMEOUT=${DEPLOY_TIMEOUT:-1200}
-echo "⏳ Waiting for deployment to be ready (timeout: ${DEPLOY_TIMEOUT}s)..."
-echo "   This may take several minutes for GPU workloads..."
-echo ""
-
-# Poll pod status every 30s while rollout is in progress
-(
-    while true; do
-        sleep 30
-        POD_STATUS=$(kubectl get pods -n "${HYPERPOD_NAMESPACE}" -l app=${PROJECT_NAME} \
-            --no-headers 2>/dev/null | head -5)
-        if [ -n "${POD_STATUS}" ]; then
-            echo "   📊 $(date +%H:%M:%S) Pod status:"
-            echo "${POD_STATUS}" | while read -r line; do echo "      ${line}"; done
-        fi
-    done
-) &
-STATUS_PID=$!
-trap "kill ${STATUS_PID} 2>/dev/null; wait ${STATUS_PID} 2>/dev/null" EXIT
-
-ROLLOUT_OUTPUT=$(kubectl rollout status deployment/${PROJECT_NAME} -n "${HYPERPOD_NAMESPACE}" --timeout=${DEPLOY_TIMEOUT}s 2>&1) || {
-    ROLLOUT_EXIT_CODE=$?
-    kill ${STATUS_PID} 2>/dev/null
-    echo ""
-    echo "❌ Deployment failed to become ready within timeout"
-    echo ""
-    echo "   Error details:"
-    echo "   ${ROLLOUT_OUTPUT}"
-    echo ""
-    echo "   Current pod state:"
-    kubectl get pods -n "${HYPERPOD_NAMESPACE}" -l app=${PROJECT_NAME} -o wide 2>/dev/null
-    echo ""
-    echo "   Debug commands:"
-    echo "   kubectl describe pods -n ${HYPERPOD_NAMESPACE} -l app=${PROJECT_NAME}"
-    echo "   kubectl logs -n ${HYPERPOD_NAMESPACE} -l app=${PROJECT_NAME} --tail=100"
-    echo ""
-    echo "   Common issues:"
-    echo "   • Image pull errors - check ECR permissions"
-    echo "   • Resource scheduling - insufficient GPU nodes"
-    echo "   • Container crash - check application logs"
-<% if (fsxVolumeHandle) { %>
-    echo "   • PVC binding errors - verify FSx CSI driver is installed on the cluster"
-    echo "     kubectl get pvc -n ${HYPERPOD_NAMESPACE}"
-    echo "     kubectl describe pvc -n ${HYPERPOD_NAMESPACE}"
-    echo "     kubectl get csidriver -o name | grep fsx"
-<% } %>
-    exit ${ROLLOUT_EXIT_CODE}
-}
-
-kill ${STATUS_PID} 2>/dev/null
-wait ${STATUS_PID} 2>/dev/null
-
-echo "✅ HyperPod EKS deployment complete!"
-echo ""
-echo "📋 Deployment Details:"
-echo "   Cluster: ${HYPERPOD_CLUSTER_NAME}"
-echo "   Namespace: ${HYPERPOD_NAMESPACE}"
-echo "   Deployment: ${PROJECT_NAME}"
-echo "   Replicas: ${HYPERPOD_REPLICAS}"
-echo "   Image: ${ECR_REPOSITORY}:${IMAGE_TAG}"
-echo ""
-echo "📋 What's next?"
-echo "   • Test your deployment:       ./do/test"
-echo "   • Check pod status:           kubectl get pods -n ${HYPERPOD_NAMESPACE}"
-echo "   • View pod logs:              kubectl logs -n ${HYPERPOD_NAMESPACE} -l app=${PROJECT_NAME}"
-<% if (typeof includeBenchmark !== 'undefined' && includeBenchmark) { %>
-echo "   • Benchmark performance:      ./do/benchmark"
-<% } %>
-echo "   • Register this deployment:   ./do/register"
-echo "   • View logs:                  ./do/logs"
-echo "   • Clean up when done:         ./do/clean hyperpod"
-
-# Write kubeconfig path to config so other scripts can use it (idempotent)
-_update_config_var() {
-    local var_name="$1" var_value="$2" config_file="${SCRIPT_DIR}/config"
-    if grep -q "^export ${var_name}=" "${config_file}" 2>/dev/null; then
-        sed -i.bak "s|^export ${var_name}=.*|export ${var_name}=\"${var_value}\"|" "${config_file}"
-        rm -f "${config_file}.bak"
-    else
-        echo "" >> "${config_file}"
-        echo "export ${var_name}=\"${var_value}\"" >> "${config_file}"
-    fi
-}
-
-_update_config_var "KUBECONFIG" "${KUBECONFIG_PATH}"
-
-<% } else if (deploymentTarget === 'batch-transform') { %>
-# ============================================================
-# SageMaker Batch Transform Deployment
-# Flow: create-model → create-transform-job → poll until completion
-# ============================================================
-
-# Source shared helpers
-source "${SCRIPT_DIR}/lib/secrets.sh"
-source "${SCRIPT_DIR}/lib/wait.sh"
-
-# Resolve container secrets (HF_TOKEN, NGC_API_KEY)
-resolve_secrets
-
-# Validate execution role ARN
-if [ -z "${ROLE_ARN:-}" ]; then
-    echo "❌ Execution role ARN not provided"
-    echo ""
-    echo "Usage:"
-    echo "  export ROLE_ARN=arn:aws:iam::ACCOUNT_ID:role/YOUR_ROLE"
-    echo "  ./do/deploy"
-    echo ""
-    echo "Or set ROLE_ARN in do/config"
-    echo ""
-    echo "The execution role must have permissions for:"
-    echo "  • SageMaker model and transform job management"
-    echo "  • ECR image access"
-    echo "  • S3 read access for input path: ${BATCH_INPUT_PATH}"
-    echo "  • S3 write access for output path: ${BATCH_OUTPUT_PATH}"
-    echo "  • CloudWatch Logs"
-    exit 3
-fi
-
-echo "   Using execution role: ${ROLE_ARN}"
-
-# Validate S3 input path
-if [ -z "${BATCH_INPUT_PATH:-}" ]; then
-    echo "❌ S3 input path not provided"
-    echo ""
-    echo "Set BATCH_INPUT_PATH in do/config or provide via CLI:"
-    echo "  export BATCH_INPUT_PATH=s3://my-bucket/input/"
-    echo "  ./do/deploy"
-    exit 3
-fi
-
-if [[ "${BATCH_INPUT_PATH}" != s3://* ]]; then
-    echo "❌ S3 input path must start with s3://"
-    echo "   Current value: ${BATCH_INPUT_PATH}"
-    echo "   Example: s3://my-bucket/input/"
-    exit 3
-fi
-
-# Validate S3 output path
-if [ -z "${BATCH_OUTPUT_PATH:-}" ]; then
-    echo "❌ S3 output path not provided"
-    echo ""
-    echo "Set BATCH_OUTPUT_PATH in do/config or provide via CLI:"
-    echo "  export BATCH_OUTPUT_PATH=s3://my-bucket/output/"
-    echo "  ./do/deploy"
-    exit 3
-fi
-
-if [[ "${BATCH_OUTPUT_PATH}" != s3://* ]]; then
-    echo "❌ S3 output path must start with s3://"
-    echo "   Current value: ${BATCH_OUTPUT_PATH}"
-    echo "   Example: s3://my-bucket/output/"
-    exit 3
-fi
-
-# ============================================================
-# Bootstrap S3 buckets for batch transform
-# ============================================================
-
-# Extract bucket names from S3 paths
-BATCH_INPUT_BUCKET=$(echo "${BATCH_INPUT_PATH}" | sed 's|s3://||' | cut -d'/' -f1)
-BATCH_OUTPUT_BUCKET=$(echo "${BATCH_OUTPUT_PATH}" | sed 's|s3://||' | cut -d'/' -f1)
-
-<% if (!batchInputPath) { %>
-# Bootstrap default S3 input bucket (check-and-create)
-echo "🔍 Checking if S3 input bucket exists: ${BATCH_INPUT_BUCKET}"
-if ! aws s3api head-bucket --bucket "${BATCH_INPUT_BUCKET}" --region "${AWS_REGION}" 2>/dev/null; then
-    echo "📦 Creating S3 input bucket: ${BATCH_INPUT_BUCKET}"
-    if [ "${AWS_REGION}" = "us-east-1" ]; then
-        if ! aws s3api create-bucket \
-            --bucket "${BATCH_INPUT_BUCKET}" \
-            --region "${AWS_REGION}"; then
-            echo "❌ Failed to create S3 input bucket: ${BATCH_INPUT_BUCKET}"
-            echo ""
-            echo "   Check that:"
-            echo "   • Your IAM credentials have s3:CreateBucket permission"
-            echo "   • The bucket name is not already taken globally"
-            exit 4
-        fi
-    else
-        if ! aws s3api create-bucket \
-            --bucket "${BATCH_INPUT_BUCKET}" \
-            --region "${AWS_REGION}" \
-            --create-bucket-configuration LocationConstraint="${AWS_REGION}"; then
-            echo "❌ Failed to create S3 input bucket: ${BATCH_INPUT_BUCKET}"
-            echo ""
-            echo "   Check that:"
-            echo "   • Your IAM credentials have s3:CreateBucket permission"
-            echo "   • The bucket name is not already taken globally"
-            exit 4
-        fi
-    fi
-    echo "✅ S3 input bucket created: ${BATCH_INPUT_BUCKET}"
-else
-    echo "✅ S3 input bucket exists: ${BATCH_INPUT_BUCKET}"
-fi
-
-# Upload sample input file if the input prefix is empty
-EXISTING_OBJECTS=$(aws s3 ls "${BATCH_INPUT_PATH}" --region "${AWS_REGION}" 2>/dev/null | head -1 || true)
-if [ -z "${EXISTING_OBJECTS}" ]; then
-    echo "📄 Uploading sample input file to ${BATCH_INPUT_PATH}"
-<% if (framework === 'transformers' && (modelServer === 'vllm' || modelServer === 'sglang')) { %>
-    echo '{"model": "<%= modelName %>", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 50}' | aws s3 cp - "${BATCH_INPUT_PATH}sample.jsonl" --region "${AWS_REGION}"
-<% } else if (framework === 'transformers') { %>
-    echo '{"inputs": "What is machine learning?", "parameters": {"max_new_tokens": 50}}' | aws s3 cp - "${BATCH_INPUT_PATH}sample.jsonl" --region "${AWS_REGION}"
-<% } else if (framework === 'diffusors') { %>
-    echo '{"prompt": "A white cat", "n": 1, "size": "512x512"}' | aws s3 cp - "${BATCH_INPUT_PATH}sample.jsonl" --region "${AWS_REGION}"
-<% } else { %>
-    echo '{"instances": [[1.0, 2.0, 3.0, 4.0]]}' | aws s3 cp - "${BATCH_INPUT_PATH}sample.jsonl" --region "${AWS_REGION}"
-<% } %>
-    echo "✅ Sample input uploaded: ${BATCH_INPUT_PATH}sample.jsonl"
-    echo "   ⚠️  Replace this with your actual input data before running production jobs"
-fi
-<% } else { %>
-# Custom S3 input path provided — skip bucket creation
-echo "✅ Using custom S3 input path: ${BATCH_INPUT_PATH}"
-<% } %>
-
-<% if (!batchOutputPath) { %>
-# Bootstrap default S3 output bucket (check-and-create, may be same as input)
-if [ "${BATCH_OUTPUT_BUCKET}" != "${BATCH_INPUT_BUCKET}" ]; then
-    echo "🔍 Checking if S3 output bucket exists: ${BATCH_OUTPUT_BUCKET}"
-    if ! aws s3api head-bucket --bucket "${BATCH_OUTPUT_BUCKET}" --region "${AWS_REGION}" 2>/dev/null; then
-        echo "📦 Creating S3 output bucket: ${BATCH_OUTPUT_BUCKET}"
-        if [ "${AWS_REGION}" = "us-east-1" ]; then
-            if ! aws s3api create-bucket \
-                --bucket "${BATCH_OUTPUT_BUCKET}" \
-                --region "${AWS_REGION}"; then
-                echo "❌ Failed to create S3 output bucket: ${BATCH_OUTPUT_BUCKET}"
-                exit 4
-            fi
-        else
-            if ! aws s3api create-bucket \
-                --bucket "${BATCH_OUTPUT_BUCKET}" \
-                --region "${AWS_REGION}" \
-                --create-bucket-configuration LocationConstraint="${AWS_REGION}"; then
-                echo "❌ Failed to create S3 output bucket: ${BATCH_OUTPUT_BUCKET}"
-                exit 4
-            fi
-        fi
-        echo "✅ S3 output bucket created: ${BATCH_OUTPUT_BUCKET}"
-    else
-        echo "✅ S3 output bucket exists: ${BATCH_OUTPUT_BUCKET}"
-    fi
-else
-    echo "✅ S3 output bucket same as input: ${BATCH_OUTPUT_BUCKET}"
-fi
-<% } else { %>
-# Custom S3 output path provided — skip bucket creation
-echo "✅ Using custom S3 output path: ${BATCH_OUTPUT_PATH}"
-<% } %>
-
-# ============================================================
-# Check for previous transform job still running
-# ============================================================
-if [ "${FORCE_NEW}" != true ] && [ -n "${TRANSFORM_JOB_NAME:-}" ]; then
-    echo "🔍 Checking previous transform job: ${TRANSFORM_JOB_NAME}"
-    PREV_JOB_STATUS=$(aws sagemaker describe-transform-job \
-        --transform-job-name "${TRANSFORM_JOB_NAME}" \
-        --region "${AWS_REGION}" \
-        --query "TransformJobStatus" \
-        --output text 2>/dev/null || echo "")
-
-    case "${PREV_JOB_STATUS}" in
-        InProgress)
-            echo "⚠️  Previous transform job is still running: ${TRANSFORM_JOB_NAME}"
-            echo "   Wait for it to complete, or stop it with:"
-            echo "   aws sagemaker stop-transform-job --transform-job-name ${TRANSFORM_JOB_NAME} --region ${AWS_REGION}"
-            echo ""
-            echo "   Use --force to create a new job anyway."
-            exit 4
-            ;;
-        Completed)
-            echo "✅ Previous transform job completed: ${TRANSFORM_JOB_NAME}"
-            echo "   Creating a new job. Results from the previous job are in:"
-            echo "   ${BATCH_OUTPUT_PATH}"
-            echo ""
-            ;;
-        *)
-            # Failed, Stopped, or not found — proceed with new job
-            ;;
-    esac
-fi
-
-# Generate unique names with timestamp
-TIMESTAMP=$(date +%s)
-MODEL_NAME_SM="${PROJECT_NAME}-batch-model-${TIMESTAMP}"
-TRANSFORM_JOB_NAME="${PROJECT_NAME}-batch-job-${TIMESTAMP}"
-
-_update_config_var "TRANSFORM_JOB_NAME" "${TRANSFORM_JOB_NAME}"
-_update_config_var "SAGEMAKER_MODEL_NAME" "${MODEL_NAME_SM}"
-
-# Step 1: Create SageMaker model
-echo "📦 Creating SageMaker model: ${MODEL_NAME_SM}"
-
-# Build primary container spec
-BATCH_PRIMARY_CONTAINER="{\"Image\":\"${ECR_REPOSITORY}:${IMAGE_TAG}\""
-if [ -n "${CONTAINER_ENV_JSON}" ]; then
-    BATCH_PRIMARY_CONTAINER="${BATCH_PRIMARY_CONTAINER},\"Environment\":{${CONTAINER_ENV_JSON}}"
-fi
-BATCH_PRIMARY_CONTAINER="${BATCH_PRIMARY_CONTAINER}}"
-
-if ! aws sagemaker create-model \
-    --model-name "${MODEL_NAME_SM}" \
-    --primary-container "${BATCH_PRIMARY_CONTAINER}" \
-    --execution-role-arn "${ROLE_ARN}" \
-    --region "${AWS_REGION}"; then
-
-    echo "❌ Failed to create SageMaker model"
-    echo "   Check that:"
-    echo "   • The execution role ARN is valid"
-    echo "   • The ECR image exists and is accessible"
-    echo "   • The IAM role has ecr:GetDownloadUrlForLayer permission"
-    exit 4
-fi
-
-echo "✅ SageMaker model created: ${MODEL_NAME_SM}"
-
-# Record model in manifest (non-blocking)
-MODEL_ARN="arn:aws:sagemaker:${AWS_REGION}:${AWS_ACCOUNT_ID}:model/${MODEL_NAME_SM}"
-./do/manifest add \
-    --type sagemaker-model \
-    --id "${MODEL_ARN}" \
-    --project "${PROJECT_NAME}" \
-    --meta "{\"modelName\":\"${MODEL_NAME_SM}\",\"region\":\"${AWS_REGION}\"}" \
-    2>/dev/null || true
-
-# Step 2: Build transform job JSON
-TRANSFORM_JOB_JSON="{
-    \"TransformJobName\": \"${TRANSFORM_JOB_NAME}\",
-    \"ModelName\": \"${MODEL_NAME_SM}\",
-    \"TransformInput\": {
-        \"DataSource\": {
-            \"S3DataSource\": {
-                \"S3DataType\": \"S3Prefix\",
-                \"S3Uri\": \"${BATCH_INPUT_PATH}\"
-            }
-        },
-        \"ContentType\": \"application/json\",
-        \"SplitType\": \"${BATCH_SPLIT_TYPE}\"
-    },
-    \"TransformOutput\": {
-        \"S3OutputPath\": \"${BATCH_OUTPUT_PATH}\"
-        $([ "${BATCH_JOIN_SOURCE:-None}" = "Input" ] && echo ",\"Accept\": \"application/json\", \"AssembleWith\": \"${BATCH_SPLIT_TYPE}\"")
-    },
-    \"TransformResources\": {
-        \"InstanceType\": \"${INSTANCE_TYPE}\",
-        \"InstanceCount\": ${BATCH_INSTANCE_COUNT}
-    },
-    \"MaxConcurrentTransforms\": ${BATCH_MAX_CONCURRENT_TRANSFORMS:-1},
-    \"MaxPayloadInMB\": ${BATCH_MAX_PAYLOAD_IN_MB:-6},
-    \"BatchStrategy\": \"${BATCH_STRATEGY}\"
-    $([ "${BATCH_JOIN_SOURCE:-None}" = "Input" ] && echo ",\"DataProcessing\": { \"JoinSource\": \"Input\" }")
-}"
-
-# Step 3: Create transform job
-echo "🚀 Creating transform job: ${TRANSFORM_JOB_NAME}"
-if ! aws sagemaker create-transform-job \
-    --cli-input-json "${TRANSFORM_JOB_JSON}" \
-    --region "${AWS_REGION}"; then
-
-    echo "❌ Failed to create transform job"
-    echo "   Check that:"
-    echo "   • The S3 input path exists and is accessible: ${BATCH_INPUT_PATH}"
-    echo "   • The S3 output path is writable: ${BATCH_OUTPUT_PATH}"
-    echo "   • The IAM role has s3:GetObject permission on the input path"
-    echo "   • The IAM role has s3:PutObject permission on the output path"
-    echo "   • The instance type is valid: ${INSTANCE_TYPE}"
-    echo "   • The instance type is available in region: ${AWS_REGION}"
-    echo "   • You have sufficient service quota for the instance type"
-    exit 4
-fi
-
-echo "✅ Transform job created: ${TRANSFORM_JOB_NAME}"
-
-# Record transform job in manifest (non-blocking)
-TRANSFORM_JOB_ARN="arn:aws:sagemaker:${AWS_REGION}:${AWS_ACCOUNT_ID}:transform-job/${TRANSFORM_JOB_NAME}"
-./do/manifest add \
-    --type sagemaker-transform-job \
-    --id "${TRANSFORM_JOB_ARN}" \
-    --project "${PROJECT_NAME}" \
-    --meta "{\"transformJobName\":\"${TRANSFORM_JOB_NAME}\",\"modelName\":\"${MODEL_NAME_SM}\",\"instanceType\":\"${INSTANCE_TYPE}\",\"region\":\"${AWS_REGION}\"}" \
-    2>/dev/null || true
-
-# Step 4: Poll transform job status until completion or failure
-echo "⏳ Waiting for transform job to complete..."
-echo "   This may take several minutes depending on dataset size..."
-echo "   If this times out, check status with:"
-echo "   aws sagemaker describe-transform-job --transform-job-name ${TRANSFORM_JOB_NAME} --region ${AWS_REGION}"
-echo ""
-
-while true; do
-    JOB_STATUS=$(aws sagemaker describe-transform-job \
-        --transform-job-name "${TRANSFORM_JOB_NAME}" \
-        --region "${AWS_REGION}" \
-        --query "TransformJobStatus" \
-        --output text 2>&1) || {
-        # Check if it was a credential expiration
-        if echo "${JOB_STATUS}" | grep -qi "expired\|token"; then
-            echo ""
-            echo "⚠️  Credentials expired, but the transform job is still running."
-            echo "   Refresh your credentials and check status with:"
-            echo "   aws sagemaker describe-transform-job --transform-job-name ${TRANSFORM_JOB_NAME} --region ${AWS_REGION} --query TransformJobStatus"
-            exit 4
-        fi
-        echo "❌ Failed to describe transform job: ${TRANSFORM_JOB_NAME}"
-        echo "   Error: ${JOB_STATUS}"
-        exit 4
-    }
-
-    case "${JOB_STATUS}" in
-        Completed)
-            echo "✅ Transform job completed successfully!"
-            break
-            ;;
-        Failed)
-            FAILURE_REASON=$(aws sagemaker describe-transform-job \
-                --transform-job-name "${TRANSFORM_JOB_NAME}" \
-                --region "${AWS_REGION}" \
-                --query "FailureReason" \
-                --output text 2>/dev/null || echo "Unknown")
-            echo "❌ Transform job failed"
-            echo "   Reason: ${FAILURE_REASON}"
-            echo ""
-            echo "   Check CloudWatch Logs for details:"
-            echo "   https://console.aws.amazon.com/cloudwatch/home?region=${AWS_REGION}#logsV2:log-groups/log-group//aws/sagemaker/TransformJobs"
-            echo ""
-            echo "   Verify that:"
-            echo "   • The S3 input path exists and contains data: ${BATCH_INPUT_PATH}"
-            echo "   • The input data format matches the container's expected format"
-            echo "   • The container's /ping and /invocations endpoints work correctly"
-            exit 4
-            ;;
-        Stopped)
-            echo "⚠️  Transform job was stopped"
-            exit 4
-            ;;
-        InProgress)
-            echo "   $(date +%H:%M:%S) Job status: InProgress..."
-            sleep 30
-            ;;
-        *)
-            echo "   $(date +%H:%M:%S) Job status: ${JOB_STATUS}..."
-            sleep 30
-            ;;
-    esac
-done
-
-echo ""
-echo "📋 Deployment Details:"
-echo "   Transform Job: ${TRANSFORM_JOB_NAME}"
-echo "   Model: ${MODEL_NAME_SM}"
-echo "   Region: ${AWS_REGION}"
-echo "   Instance Type: ${INSTANCE_TYPE}"
-echo "   Instance Count: ${BATCH_INSTANCE_COUNT}"
-echo "   Image: ${ECR_REPOSITORY}:${IMAGE_TAG}"
-echo "   S3 Input: ${BATCH_INPUT_PATH}"
-echo "   S3 Output: ${BATCH_OUTPUT_PATH}"
-echo "   Split Type: ${BATCH_SPLIT_TYPE}"
-echo "   Strategy: ${BATCH_STRATEGY}"
-echo ""
-
-# Download results locally
-LOCAL_OUTPUT_DIR="${SCRIPT_DIR}/../batch-output"
-mkdir -p "${LOCAL_OUTPUT_DIR}"
-echo "📥 Downloading results to ${LOCAL_OUTPUT_DIR}/"
-if aws s3 sync "${BATCH_OUTPUT_PATH}" "${LOCAL_OUTPUT_DIR}/" --region "${AWS_REGION}"; then
-    DOWNLOADED=$(ls -1 "${LOCAL_OUTPUT_DIR}" 2>/dev/null | wc -l | tr -d ' ')
-    echo "✅ Downloaded ${DOWNLOADED} file(s) to ${LOCAL_OUTPUT_DIR}/"
-    echo ""
-
-    # Display first output file preview
-    FIRST_FILE=$(ls -1 "${LOCAL_OUTPUT_DIR}" 2>/dev/null | head -1)
-    if [ -n "${FIRST_FILE}" ]; then
-        echo "📄 Sample output (${FIRST_FILE}):"
-        head -5 "${LOCAL_OUTPUT_DIR}/${FIRST_FILE}"
-        LINES=$(wc -l < "${LOCAL_OUTPUT_DIR}/${FIRST_FILE}" | tr -d ' ')
-        if [ "${LINES}" -gt 5 ]; then
-            echo "   ... (${LINES} total lines)"
-        fi
-    fi
-else
-    echo "⚠️  Could not download output files"
-fi
-
-echo ""
-echo "📋 What's next?"
-echo "   • View results:               cat batch-output/"
-echo "   • Review results:             ./do/test"
-echo "   • Register this deployment:   ./do/register"
-echo "   • View logs:                  ./do/logs"
-echo "   • Clean up when done:         ./do/clean"
-
-<% } %>
+<%- include('deploy.d/' + (deploymentTarget === 'realtime-inference' ? 'managed-inference' : deploymentTarget)) %>