@aws/ml-container-creator 0.3.0 → 0.5.0

package/templates/do/config

@@ -10,6 +10,11 @@ export DEPLOYMENT_CONFIG="<%= deploymentConfig %>"
 export FRAMEWORK="<%= framework %>"
 export MODEL_SERVER="<%= modelServer %>"
 
+<% if (typeof enableLora !== 'undefined' && enableLora) { %>
+# LoRA adapter serving
+export ENABLE_LORA=true
+<% } %>
+
 # AWS configuration
 export AWS_REGION="<%= awsRegion %>"
 export ECR_REPOSITORY_NAME="ml-container-creator"
@@ -26,10 +31,26 @@ export DEPLOYMENT_TARGET="<%= deploymentTarget %>"
 
 <% if (deploymentTarget === 'realtime-inference') { %>
 # SageMaker Real-Time Inference configuration
+<% if (typeof existingEndpointName !== 'undefined' && existingEndpointName) { %>
+# External endpoint — attaching IC to an existing running endpoint
+export ENDPOINT_NAME="<%= existingEndpointName %>"
+export ENDPOINT_EXTERNAL=true
+<% } else { %>
 export INSTANCE_TYPE="<%= instanceType %>"
+<% if (typeof instancePools !== 'undefined' && instancePools && instancePools.length > 1) { %>
+# Instance pools: heterogeneous instance types with priority-based fallback
+# Priority = selection order (1 = preferred, higher = fallback)
+export INSTANCE_POOLS='<%= JSON.stringify(instancePools) %>'
+<% } %>
 <% if (inferenceAmiVersion) { %>
 export INFERENCE_AMI_VERSION="<%= inferenceAmiVersion %>"
 <% } %>
+<% if (typeof capacityReservationArn !== 'undefined' && capacityReservationArn) { %>
+# Note: Capacity reservations and instance pools (INSTANCE_POOLS) are mutually exclusive.
+# If both are set, the capacity reservation takes precedence and INSTANCE_POOLS is ignored.
+export CAPACITY_RESERVATION_ARN="<%= capacityReservationArn %>"
+<% } %>
+<% } %>
 <% } %>
 
 <% if (deploymentTarget === 'async-inference') { %>
@@ -126,6 +147,8 @@ export IC_MEMORY_SIZE="<%= icMemorySize %>"
 <% } %>
 <% if (typeof icGpuCount !== 'undefined' && icGpuCount != null) { %>
 export IC_GPU_COUNT="<%= icGpuCount %>"
+<% } else { %>
+export IC_GPU_COUNT="${IC_GPU_COUNT:-1}"
 <% } %>
 <% if (typeof icCopyCount !== 'undefined' && icCopyCount != null) { %>
 export IC_COPY_COUNT="<%= icCopyCount %>"
@@ -186,6 +209,26 @@ export MODEL_FORMAT="<%= modelFormat %>"
 export ROLE_ARN="<%= roleArn %>"
 <% } %>
 
+<% if (typeof includeBenchmark !== 'undefined' && includeBenchmark) { %>
+# SageMaker AI Benchmarking configuration
+export BENCHMARK_CONCURRENCY="<%= benchmarkConcurrency %>"
+export BENCHMARK_INPUT_TOKENS_MEAN="<%= benchmarkInputTokensMean %>"
+export BENCHMARK_OUTPUT_TOKENS_MEAN="<%= benchmarkOutputTokensMean %>"
+export BENCHMARK_STREAMING="<%= benchmarkStreaming %>"
+<% if (benchmarkRequestCount) { %>
+export BENCHMARK_REQUEST_COUNT="<%= benchmarkRequestCount %>"
+<% } else { %>
+export BENCHMARK_REQUEST_COUNT=""
+<% } %>
+<% if (benchmarkS3OutputPath) { %>
+export BENCHMARK_S3_OUTPUT_PATH="<%= benchmarkS3OutputPath %>"
+<% } else { %>
+export BENCHMARK_S3_OUTPUT_PATH="s3://ml-container-creator-benchmark-${AWS_REGION}-$(aws sts get-caller-identity --query Account --output text)/${PROJECT_NAME}/"
+<% } %>
+export BENCHMARK_JOB_NAME=""
+export BENCHMARK_WORKLOAD_CONFIG_NAME=""
+<% } %>
+
 <% if (orderedEnvVars && orderedEnvVars.length > 0) { %>
 # Runtime environment variables (from catalog)
 <% orderedEnvVars.forEach(({ key, value }) => { %>
@@ -193,13 +236,11 @@ export <%= key %>=${<%= key %>:-<%= value %>}
 <% }); %>
 <% } %>
 
-<% if (baseImage) { %>
-export BASE_IMAGE=${BASE_IMAGE:-<%= baseImage %>}
-<% } %>
+export BASE_IMAGE=${BASE_IMAGE:-<%= baseImage || '' %>}
 
 # Allow environment variable overrides
 export AWS_REGION=${AWS_REGION:-<%= awsRegion %>}
-<% if (deploymentTarget === 'realtime-inference' || deploymentTarget === 'async-inference' || deploymentTarget === 'batch-transform') { %>
+<% if ((deploymentTarget === 'realtime-inference' && !(typeof existingEndpointName !== 'undefined' && existingEndpointName)) || deploymentTarget === 'async-inference' || deploymentTarget === 'batch-transform') { %>
 export INSTANCE_TYPE=${INSTANCE_TYPE:-<%= instanceType %>}
 <% } %>
 export ECR_REPOSITORY_NAME=${ECR_REPOSITORY_NAME:-ml-container-creator}
@@ -254,7 +295,11 @@ echo "   Model env vars: <%= Object.keys(modelEnvVars).length %>"
 echo "   Server env vars: <%= Object.keys(serverEnvVars).length %>"
 <% } %>
 <% if (deploymentTarget === 'realtime-inference') { %>
+<% if (typeof existingEndpointName !== 'undefined' && existingEndpointName) { %>
+echo "   Endpoint: ${ENDPOINT_NAME} (external)"
+<% } else { %>
 echo "   Instance: ${INSTANCE_TYPE}"
+<% } %>
 <% } else if (deploymentTarget === 'async-inference') { %>
 echo "   Instance: ${INSTANCE_TYPE}"
 echo "   S3 output: ${ASYNC_S3_OUTPUT_PATH}"