@aws/ml-container-creator 0.13.5 → 0.15.0

package/templates/do/benchmark

@@ -39,17 +39,18 @@ while [ $# -gt 0 ]; do
             echo "Options:"
             echo "  --status            Check job status; if completed, download results + write to Athena"
             echo "  --ic <name>         Benchmark a specific inference component"
-            echo "  --adapter <name>    Benchmark a specific LoRA adapter IC"
+            echo "  --adapter <name>    Benchmark a specific LoRA adapter (routes via adapter IC, records adapter_name in Athena)"
             echo "  --force             Create a new benchmark job even if one is already running"
             echo "  --clean             Delete workload config and benchmark job after displaying results"
             echo "  --no-stale-warning  Suppress schema registry staleness warning"
-            echo "  --no-stale-warning  Suppress schema registry staleness warning"
             echo ""
             echo "IC resolution:"
             echo "  --adapter <name> Use ADAPTER_IC_NAME from do/adapters/<name>.conf"
             echo "  --ic <name>      Use IC_DEPLOYED_NAME from do/ic/<name>.conf"
             echo "  (no flag)        Use first IC in do/ic/ alphabetically, or legacy config"
             echo ""
+            echo "Adapter benchmarks are differentiated from base model runs in Athena via the adapter_name column."
+            echo ""
             echo "Status:"
             echo "  After interrupting a running benchmark, use --status to check completion"
             echo "  and trigger results download + Athena write."
@@ -113,10 +114,17 @@ if [ "${ARG_STATUS}" = true ]; then
                     aws s3 sync "${RESULTS_S3_PATH}" "${LOCAL_RESULTS_DIR}/output/" \
                         --region "${AWS_REGION}" --quiet
                     # Untar if output.tar.gz exists
-                    local tar_file
+                    tar_file=""
                     tar_file=$(find "${LOCAL_RESULTS_DIR}" -name "output.tar.gz" -type f 2>/dev/null | head -1)
                     if [ -n "${tar_file}" ]; then
-                        tar -xzf "${tar_file}" --strip-components=1 -C "${LOCAL_RESULTS_DIR}/output/" 2>/dev/null || true
+                        # Detect whether tar has a leading directory prefix
+                        _tar_first=""
+                        _tar_first=$(tar -tzf "${tar_file}" 2>/dev/null | head -1)
+                        if echo "${_tar_first}" | grep -qE '^[^/]+/$'; then
+                            tar -xzf "${tar_file}" --strip-components=1 -C "${LOCAL_RESULTS_DIR}/output/" 2>/dev/null || true
+                        else
+                            tar -xzf "${tar_file}" -C "${LOCAL_RESULTS_DIR}/output/" 2>/dev/null || true
+                        fi
                     fi
                     # Re-search after extraction
                     RESULTS_JSONL=$(find "${LOCAL_RESULTS_DIR}" -name "profile_export.jsonl" -type f 2>/dev/null | head -1)
@@ -145,7 +153,8 @@ if [ "${ARG_STATUS}" = true ]; then
                         --workload "${BENCHMARK_WORKLOAD:-manual}" \
                         --concurrency "${BENCHMARK_CONCURRENCY:-2}" \
                         --bucket "${CI_BENCHMARK_RESULTS_BUCKET}" \
-                        --region "${AWS_REGION:-${REGION}}"; then
+                        --region "${AWS_REGION:-${REGION}}" \
+                        ${ADAPTER_ARG:+--adapter-name "${ADAPTER_ARG}"}; then
                         echo "   ✅ Results persisted to Athena"
                     else
                         echo "   ⚠️  Athena write failed (non-fatal)"
@@ -549,7 +558,8 @@ print(f'Combined {n_metrics} concurrency level results')
                 --project-name "${PROJECT_NAME}" \
                 --workload "${BENCHMARK_WORKLOAD:-manual}" \
                 --bucket "${CI_BENCHMARK_RESULTS_BUCKET}" \
-                --region "${AWS_REGION:-${REGION}}"; then
+                --region "${AWS_REGION:-${REGION}}" \
+                ${ADAPTER_ARG:+--adapter-name "${ADAPTER_ARG}"}; then
                 echo "✅ Multi-level benchmark results persisted to S3"
             else
                 echo "⚠️  Failed to persist multi-level benchmark results to Athena (non-fatal)"
@@ -1087,7 +1097,16 @@ if [ "${JOB_STATUS}" = "Completed" ]; then
             # Extract any tar.gz archives (benchmark service packages results as output.tar.gz)
             for ARCHIVE in $(find "${LOCAL_RESULTS_DIR}" -name "*.tar.gz" -type f 2>/dev/null); do
                 ARCHIVE_DIR=$(dirname "${ARCHIVE}")
-                tar -xzf "${ARCHIVE}" --strip-components=1 -C "${ARCHIVE_DIR}" 2>/dev/null || true
+                # Detect whether tar has a leading directory prefix to strip.
+                # Some AIPerf versions wrap in output/, others are flat.
+                _TAR_FIRST=$(tar -tzf "${ARCHIVE}" 2>/dev/null | head -1)
+                if echo "${_TAR_FIRST}" | grep -qE '^[^/]+/$'; then
+                    # Leading directory (e.g., "output/") — strip it
+                    tar -xzf "${ARCHIVE}" --strip-components=1 -C "${ARCHIVE_DIR}" 2>/dev/null || true
+                else
+                    # Flat archive — extract as-is
+                    tar -xzf "${ARCHIVE}" -C "${ARCHIVE_DIR}" 2>/dev/null || true
+                fi
             done
 
             # Look for specific result files (priority: JSONL > aiperf JSON)
@@ -1346,7 +1365,8 @@ except Exception as e:
             --workload "${BENCHMARK_WORKLOAD:-manual}" \
             --concurrency "${BENCHMARK_CONCURRENCY}" \
             --bucket "${CI_BENCHMARK_RESULTS_BUCKET}" \
-            --region "${AWS_REGION:-${REGION}}"; then
+            --region "${AWS_REGION:-${REGION}}" \
+            ${ADAPTER_ARG:+--adapter-name "${ADAPTER_ARG}"}; then
             echo "✅ Benchmark results persisted to S3"
         else
             echo "⚠️  Failed to persist benchmark results to Athena (non-fatal)"

package/templates/do/config

@@ -211,6 +211,26 @@ export <%= key %>=${<%= key %>:-<%= value %>}
 <% }); %>
 <% } %>
 
+<% if (typeof icEnvVars !== 'undefined' && icEnvVars && Object.keys(icEnvVars).length > 0) { %>
+# Deploy-time IC environment variables (applied at IC creation via SDK v3, not baked into image)
+# Max 16 vars, max 1024 chars per key/value
+# WARNING: Do not store raw secrets here. Use Secrets Manager ARN pattern instead:
+#   export IC_ENV_HF_TOKEN_ARN=arn:aws:secretsmanager:REGION:ACCOUNT:secret:NAME
+#   Runtime code resolves the ARN to the secret value (same pattern as HF_TOKEN_ARN in do/build).
+<% Object.entries(icEnvVars).forEach(([key, value]) => { %>
+export IC_ENV_<%= key %>=${IC_ENV_<%= key %>:-<%= value %>}
+<% }); %>
+<% } else if (deploymentTarget === 'realtime-inference') { %>
+# ─── Deploy-time IC environment variables (uncomment to configure) ─────────────
+# These are passed as the Environment field in InferenceComponent.create() at deploy time.
+# They do NOT affect the Docker build — build-time env vars remain in Dockerfile via ENV.
+# Max 16 vars, max 1024 chars per key/value.
+# WARNING: Do not store raw secrets here. Use Secrets Manager ARN pattern instead:
+#   export IC_ENV_HF_TOKEN_ARN=arn:aws:secretsmanager:REGION:ACCOUNT:secret:NAME
+# export IC_ENV_VLLM_MAX_MODEL_LEN=8192
+# export IC_ENV_VLLM_GPU_MEMORY_UTILIZATION=0.85
+<% } %>
+
 # Framework-specific configuration
 <% if (framework === 'transformers') { %>
 export MODEL_NAME="<%= modelName %>"

package/templates/do/lib/inference-component.sh

@@ -6,6 +6,50 @@
 #   PROJECT_NAME, ENDPOINT_NAME, ECR_REPOSITORY, AWS_REGION, CONTAINER_ENV_JSON
 # Also expects _update_config_var() to be available (from wait.sh).
 
+# _collect_ic_env_vars()
+#   Reads IC_ENV_* prefixed variables from the environment (sourced from do/config),
+#   strips the IC_ENV_ prefix, validates constraints, and outputs JSON key-value pairs.
+#   Constraints: max 16 entries, max 1024 chars per key/value.
+#   IC_ENV_* overrides take precedence over CONTAINER_ENV_JSON.
+#
+#   Sets IC_ENV_OVERRIDE in the caller's scope.
+_collect_ic_env_vars() {
+    IC_ENV_OVERRIDE=""
+    local ic_env_count=0
+
+    while IFS='=' read -r full_key value; do
+        # Skip empty lines
+        [ -z "${full_key}" ] && continue
+
+        local stripped_key="${full_key#IC_ENV_}"
+
+        # Validate key length (AC-3.4)
+        if [ ${#stripped_key} -gt 1024 ]; then
+            echo "⚠️  IC_ENV_${stripped_key}: key exceeds 1024 chars, skipping" >&2
+            continue
+        fi
+
+        # Validate value length (AC-3.4)
+        if [ ${#value} -gt 1024 ]; then
+            echo "⚠️  IC_ENV_${stripped_key}: value exceeds 1024 chars, skipping" >&2
+            continue
+        fi
+
+        ic_env_count=$((ic_env_count + 1))
+
+        # Max 16 env vars (AC-3.3)
+        if [ ${ic_env_count} -gt 16 ]; then
+            echo "⚠️  More than 16 IC_ENV_* variables defined. Using first 16 only." >&2
+            break
+        fi
+
+        if [ -n "${IC_ENV_OVERRIDE}" ]; then
+            IC_ENV_OVERRIDE="${IC_ENV_OVERRIDE},"
+        fi
+        IC_ENV_OVERRIDE="${IC_ENV_OVERRIDE}\"${stripped_key}\":\"${value}\""
+    done < <(env | grep "^IC_ENV_" | sort)
+}
+
 # create_inference_component <ic_config_file>
 #   Creates an inference component from a per-IC config file.
 #
@@ -17,6 +61,10 @@
 #     IC_STARTUP_TIMEOUT — container startup health check timeout in seconds (default: 900)
 #     IC_CONTAINER_ENV_EXTRA — optional extra env vars in "KEY":"value" format
 #
+#   IC_ENV_* prefixed vars from do/config are collected, validated, and passed
+#   as the Environment field in InferenceComponent.create() via SDK v3.
+#   Precedence: IC_ENV_* > IC_CONTAINER_ENV_EXTRA > CONTAINER_ENV_JSON
+#
 #   Multi-spec support (for heterogeneous instance pools):
 #     IC_MULTI_SPEC      — set to "true" to use Specifications (plural) array
 #     IC_SPEC_COUNT      — number of spec entries (e.g., 2)
@@ -38,6 +86,9 @@ create_inference_component() {
     # Source the IC config to get per-IC settings
     source "${ic_conf}"
 
+    # Collect IC_ENV_* overrides from environment (sourced from do/config)
+    _collect_ic_env_vars
+
     local ic_timestamp
     ic_timestamp=$(date +%s)
     local ic_basename
@@ -48,9 +99,11 @@ create_inference_component() {
     local container_spec="{\"Image\":\"${ECR_REPOSITORY}:${IC_IMAGE_TAG:-${PROJECT_NAME}-latest}\""
     # Always inject IC name for CW log forwarder
     local ic_env="\"INFERENCE_COMPONENT_NAME\":\"${ic_name}\""
-    if [ -n "${CONTAINER_ENV_JSON}${IC_CONTAINER_ENV_EXTRA:-}" ]; then
-        local env_json="${CONTAINER_ENV_JSON}"
-        [ -n "${IC_CONTAINER_ENV_EXTRA:-}" ] && env_json="${env_json:+${env_json},}${IC_CONTAINER_ENV_EXTRA}"
+    # Build environment JSON with precedence: IC_ENV_* > IC_CONTAINER_ENV_EXTRA > CONTAINER_ENV_JSON
+    local env_json="${CONTAINER_ENV_JSON}"
+    [ -n "${IC_CONTAINER_ENV_EXTRA:-}" ] && env_json="${env_json:+${env_json},}${IC_CONTAINER_ENV_EXTRA}"
+    [ -n "${IC_ENV_OVERRIDE:-}" ] && env_json="${env_json:+${env_json},}${IC_ENV_OVERRIDE}"
+    if [ -n "${env_json}" ]; then
         container_spec="${container_spec},\"Environment\":{${ic_env},${env_json}}"
     else
         container_spec="${container_spec},\"Environment\":{${ic_env}}"