feat(worker): add offline capacity priors for ModelDeploymentScaler

.agents/skills/gpustack-operator-benchmark/cases/case-5-autoscaling.sh

@@ -1,6 +1,8 @@
 #!/usr/bin/env bash
 #
-# CASE 5 — ModelDeployment autoscaling benchmark, two waves over 8 GPUs   (MUTATING, self-recovering)
+# CASE 5: capacity comparison or legacy topology benchmark (mutating).
+# Capacity comparison: --capacity-comparison --config FILE --out DIR --validate-only|--execute
+# The description below applies to the legacy topology mode.
 #
 #   case-5-autoscaling.sh [OUT_DIR] [--parallel|--serial] [--arm <arm-name>]
 #
@@ -56,6 +58,13 @@
 set -uo pipefail
 
 CASE_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+if [[ "${1:-}" == "--capacity-comparison" ]]; then
+  shift
+  exec bash "${CASE_DIR}/../scripts/autoscaling_rounds.sh" "$@"
+fi
+if [[ "${1:-}" == "--legacy-topology" ]]; then
+  shift
+fi
 SKILL_DIR="$(cd "${CASE_DIR}/.." && pwd)"
 REPO="$(cd "${SKILL_DIR}/../../.." && pwd)"
 export PATH="${REPO}/.agents/skills/_e2e-lib/scripts/kubectl-shim:${PATH}"
@@ -174,7 +183,22 @@ arm_kind() {
 arm_roles() {
   case "$1" in
     vllm-epd-predictive)
-      cat <<'ROLES'
+      local cap
+      cap="$(kubectl get instancetype.worker.gpustack.ai -o jsonpath='{.items[?(@.spec.acceleratable==true)].status.accelerator.capacity}' 2>/dev/null | awk '{print $1}')"
+      if [ "${cap:-8}" -lt 6 ] 2>/dev/null; then
+        cat <<'ROLES'
+    - name: encode
+      minReplicas: 1
+      maxReplicas: 1
+    - name: prefill
+      minReplicas: 1
+      maxReplicas: 2
+    - name: decode
+      minReplicas: 1
+      maxReplicas: 1
+ROLES
+      else
+        cat <<'ROLES'
     - name: encode
       minReplicas: 1
       maxReplicas: 2
@@ -185,6 +209,7 @@ arm_roles() {
       minReplicas: 1
       maxReplicas: 2
 ROLES
+      fi
       ;;
     *)
       cat <<'ROLES'
@@ -349,7 +374,14 @@ deploy_arm() { # <manifest> <deploy>
 # in the file is rewritten on the way in, because kubectl refuses an object whose namespace
 # differs from the one it was told to use.
 apply_manifest() { # <manifest>
-  sed "s/^  namespace: gpustack-bench$/  namespace: ${NS}/" "${SKILL_DIR}/manifests/$1" | kubectl apply -f - >/dev/null 2>&1
+  local inst_type="${BENCH_QUAL_INSTANCE_TYPE:-bench-gpu}"
+  if ! kubectl get instancetype.worker.gpustack.ai "$inst_type" >/dev/null 2>&1; then
+    inst_type="$(kubectl get instancetype.worker.gpustack.ai -o jsonpath='{.items[?(@.spec.acceleratable==true)].metadata.name}' 2>/dev/null | awk '{print $1}')"
+    inst_type="${inst_type:-bench-gpu}"
+  fi
+  sed -e "s/^  namespace: gpustack-bench$/  namespace: ${NS}/" \
+      -e "s/^      instanceType: bench-gpu$/      instanceType: ${inst_type}/" \
+      "${SKILL_DIR}/manifests/$1" | kubectl apply -f - >/dev/null 2>&1
 }
 
 drop_arm() { # <deploy> [scaler]
@@ -750,7 +782,7 @@ write_summary() {
 # above it, because a summary that cannot see a broken case is not a summary of the run.
 write_summary_md() {
   {
-    echo "# Autoscaling benchmark"
+    echo "# Autoscaling benchmark: legacy topology"
     echo
     echo "Generated at $(date -u +%Y-%m-%dT%H:%M:%SZ), mode ${MODE}."
     echo
@@ -853,7 +885,7 @@ jq -n \
       jq -c --arg arm "$arm" '{key: $arm, value: .}' "${OUT_BASE}/${arm}/summary.json" 2>/dev/null
     fi
   done | jq -s 'from_entries' 2>/dev/null || echo '{}')" \
-  '{generated_at: $generated_at, mode: $mode, e2e: {case128: $case128, case129: $case129}, arms: $arms}' \
+  '{comparisonKind: "legacy-topology", generated_at: $generated_at, mode: $mode, e2e: {case128: $case128, case129: $case129}, arms: $arms}' \
   > "${OUT_BASE}/comparison.json"
 
 # The summary a reader opens first: one row per arm and phase, and the exit code of each E2E

.agents/skills/gpustack-operator-benchmark/cases/case-5-hardware-qualification.sh

@@ -0,0 +1,11 @@
+#!/usr/bin/env bash
+# Hardware qualification for the ModelDeployment scaling chain on the 8x NVIDIA H200 SXM node.
+# Modes: --prepare, or --suite {sweep,capacity-arms,sim-sensitivity,builtin-catalog,pd-epd};
+# every case is bounded and dumps failure_summary.json when it fails or overruns.
+set -euo pipefail
+CASE_DIR="$(cd "$(dirname "${BASH_SOURCE[0]}")" && pwd)"
+SKILL_DIR="$(cd "${CASE_DIR}/.." && pwd)"
+REPO="$(cd "${SKILL_DIR}/../../.." && pwd)"
+export PATH="${REPO}/.agents/skills/_e2e-lib/scripts/kubectl-shim:${PATH}"
+export BENCH_INCLUSTER=1
+exec python3 -B "${SKILL_DIR}/scripts/autoscaling_qualification.py" "$@"

.agents/skills/gpustack-operator-benchmark/manifests/autoscaling-comparison.json

@@ -0,0 +1,43 @@
+{
+  "schemaVersion": "1",
+  "context": "REPLACE_WITH_EXPLICIT_CONTEXT",
+  "namespace": "capacity-comparison",
+  "namePrefix": "capacity",
+  "deploymentFile": "qualified-deployment.json",
+  "scalerFiles": {
+    "frozen-manual": "qualified-frozen-scaler.json",
+    "tuned-manual": "qualified-tuned-scaler.json",
+    "catalog": "qualified-catalog-scaler.json"
+  },
+  "catalogFile": "qualified-local-catalog-configmap.json",
+  "fitRunIDs": ["REPLACE_WITH_OFFLINE_FIT_RUN_ID"],
+  "evaluationTraceID": "REPLACE_WITH_SEPARATE_EVALUATION_ID",
+  "clusterQueues": ["REPLACE_WITH_DEDICATED_CLUSTER_QUEUE"],
+  "repeats": 3,
+  "order": "alternating",
+  "seed": 71,
+  "startCondition": "operator-declared-warm-cache/fresh-pods",
+  "driver": {
+    "namespace": "benchmark-driver",
+    "pod": "bench-driver",
+    "container": "driver"
+  },
+  "workload": {
+    "model": "REPLACE_WITH_QUALIFIED_MODEL",
+    "inputTokens": 512,
+    "outputTokens": 64,
+    "tokenOffset": 0,
+    "requestTimeoutSeconds": 60,
+    "phases": [
+      {"name": "ramp", "durationSeconds": 60, "concurrency": 4},
+      {"name": "sustain", "durationSeconds": 120, "concurrency": 8},
+      {"name": "cooldown", "durationSeconds": 600, "concurrency": 0}
+    ]
+  },
+  "sampleIntervalSeconds": 2,
+  "operationTimeoutSeconds": 30,
+  "readinessTimeoutSeconds": 600,
+  "cleanupTimeoutSeconds": 300,
+  "startupAllowanceSeconds": 120,
+  "retirementAllowanceSeconds": 120
+}