ledger(H7x): REFUTED on the Xeon 6741P; OVMS decodes 1.22x EAT 1.1.0's vLLM CPU engine at matched 16-bit weights

bench/xhost/results/2026-10-21/H7x-smokes/smoke7/chain.log

@@ -0,0 +1,4 @@
+== H7x-smoke7 waiting for the lock 2026-11-01T14:31:13Z
+== H7x-smoke7 lock taken 2026-11-01T14:31:13Z
+== H7x-smoke7 start: env H7X_SMOKE=1 H7X_ORDER=B C H7X_READY_S=1800 H7X_VLLM_MEM=fit K3D_GPU_IMAGE=cryptoandcoffee/k3d-gpu:0.1.61 K3D_GPU_VENDORS=none K3D_GPU_K3S_ARGS=--kubelet-arg=eviction-hard=memory.available<100Mi,nodefs.available<1%,imagefs.available<1%,nodefs.inodesFree<1% K3D_GPU_CLUSTER=k3dgpu-x-h7 H7X_CPUS=16 H7X_NODE_CPUS=0-7,48-55 APB_MODELS=/home/devcloud/k3dgpu-xhost/models APB_EAT_CHART=/home/devcloud/k3dgpu-xhost/eat-src/core/helm-charts/vllm APB_OUT=/home/devcloud/k3dgpu-xhost/runs/H7x-smoke7 bash /home/devcloud/k3dgpu-xhost/k3d-gpu-h7x/bench/apb/h7x.sh 2026-11-01T14:31:13Z
+== H7x-smoke7 rc=0; clusters deleted: none; now:  2026-11-01T14:36:33Z

bench/xhost/results/2026-10-21/H7x-smokes/smoke7/h7x-20261101-143113/agent.err

@@ -0,0 +1 @@
+==> agent: OpenVINO/Qwen3-4B-fp16-ov as 'Qwen3-4B-fp16-ov' on CPU, 16 CPU(s)

bench/xhost/results/2026-10-21/H7x-smokes/smoke7/h7x-20261101-143113/agent.yaml

@@ -0,0 +1,104 @@
+apiVersion: v1
+kind: Namespace
+metadata:
+  name: k3d-gpu-agent
+---
+# Models and the compiled-model cache live here, so restarts never re-download.
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: ovms-models
+  namespace: k3d-gpu-agent
+spec:
+  accessModes: [ReadWriteOnce]
+  resources:
+    requests:
+      storage: 20Gi
+---
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: ovms
+  namespace: k3d-gpu-agent
+  labels:
+    app: ovms
+spec:
+  replicas: 1
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      app: ovms
+  template:
+    metadata:
+      labels:
+        app: ovms
+      annotations:
+        k3d-gpu.io/agent-device: CPU
+    spec:
+      securityContext:
+        runAsUser: 5000
+        runAsGroup: 5000
+        fsGroup: 5000
+      initContainers:
+        # --pull downloads the model once and writes its graph.pbtxt (device,
+        # parsers). On later starts the files are already there and only the
+        # graph is rewritten, so a device change takes effect with no download.
+        - name: pull
+          image: openvino/model_server:2026.4.1-gpu@sha256:a928903fe429f1e3e27107cc45932d30d95febfd2d1159dc50869c3eac76fc58
+          imagePullPolicy: IfNotPresent
+          args:
+            - --pull
+            - --source_model=OpenVINO/Qwen3-4B-fp16-ov
+            - --model_repository_path=/models
+            - --model_name=Qwen3-4B-fp16-ov
+            - --task=text_generation
+            - --target_device=CPU
+            - --tool_parser=hermes3
+            - --reasoning_parser=qwen3
+          volumeMounts:
+            - name: models
+              mountPath: /models
+      containers:
+        - name: ovms
+          image: openvino/model_server:2026.4.1-gpu@sha256:a928903fe429f1e3e27107cc45932d30d95febfd2d1159dc50869c3eac76fc58
+          imagePullPolicy: IfNotPresent
+          args:
+            - --rest_port=8000
+            - --model_path=/models/OpenVINO/Qwen3-4B-fp16-ov
+            - --model_name=Qwen3-4B-fp16-ov
+            - --cache_dir=/models/.ov_cache
+            - --log_level=INFO
+          ports:
+            - name: http
+              containerPort: 8000
+          # No liveness probe: compiling a model for a GPU takes minutes.
+          readinessProbe:
+            httpGet:
+              path: /v2/health/ready
+              port: http
+            periodSeconds: 5
+          resources:
+            requests:
+              cpu: "16"
+              memory: 6Gi
+          volumeMounts:
+            - name: models
+              mountPath: /models
+      volumes:
+        - name: models
+          persistentVolumeClaim:
+            claimName: ovms-models
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: ovms
+  namespace: k3d-gpu-agent
+spec:
+  selector:
+    app: ovms
+  ports:
+    - name: http
+      port: 8000
+      targetPort: http

bench/xhost/results/2026-10-21/H7x-smokes/smoke7/h7x-20261101-143113/chart-check.txt

@@ -0,0 +1,3 @@
+xeon-values.yaml: OK
+values.yaml: OK
+templates/deployment.yaml: OK

bench/xhost/results/2026-10-21/H7x-smokes/smoke7/h7x-20261101-143113/chart.sha256

@@ -0,0 +1,3 @@
+05c1e55b0864e6b576cc84e7ba330f45021f855e84f677f4e68add24d15a4da0  xeon-values.yaml
+636ff3968e4983cdd9cfbb9372220674921a4c725483a7744527842370e9182e  values.yaml
+93ef991af65dc4f448b504b9934131ee9acb59135962bab6103faf35ab1a226d  templates/deployment.yaml

bench/xhost/results/2026-10-21/H7x-smokes/smoke7/h7x-20261101-143113/commit.txt

@@ -0,0 +1 @@
+99789454a376979294f151a44129ac140bde4952

bench/xhost/results/2026-10-21/H7x-smokes/smoke7/h7x-20261101-143113/cpus.txt

@@ -0,0 +1,2 @@
+B Cpus_allowed_list:	0-7,48-55
+C Cpus_allowed_list:	0-7,48-55

bench/xhost/results/2026-10-21/H7x-smokes/smoke7/h7x-20261101-143113/eat.yaml

@@ -0,0 +1,252 @@
+---
+# Source: vllm/templates/secret.yaml
+# Copyright (C) 2025-2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: v1
+kind: Secret
+metadata:
+  name: vllm-qwen-3-4b-cpu-secret
+  labels:
+    app:
+    helm.sh/chart: vllm-0-latest
+    app.kubernetes.io/name: vllm
+    app.kubernetes.io/instance: vllm-qwen-3-4b-cpu
+    app.kubernetes.io/version: "0.5"
+    app.kubernetes.io/managed-by: Helm
+type: Opaque
+data:
+---
+# Source: vllm/templates/configmap.yaml
+# Copyright (C) 2025-2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: vllm-qwen-3-4b-cpu-config
+  labels:
+    helm.sh/chart: vllm-0-latest
+    app.kubernetes.io/name: vllm
+    app.kubernetes.io/instance: vllm-qwen-3-4b-cpu
+    app.kubernetes.io/version: "0.5"
+    app.kubernetes.io/managed-by: Helm
+data:
+  http_proxy: ""
+  https_proxy: ""
+  no_proxy: ""
+  TORCHINDUCTOR_CACHE_DIR: "/tmp"
+  NUMBA_CACHE_DIR: "/tmp"
+  HF_HOME: "/data"
+  OUTLINES_CACHE_DIR: "/tmp/.cache/outlines"
+  VLLM_NO_USAGE_STATS: "1"
+  DO_NOT_TRACK: "1"
+  HF_HUB_DISABLE_XET: "1"
+  TORCH_INDUCTOR_CACHE_DIR: "/tmp/torchinductor_cache"
+  TRITON_CACHE_DIR: "/tmp/.triton"
+  USER: "vllm"
+  VLLM_ALLOW_LONG_MAX_MODEL_LEN: "1"
+  VLLM_CPU_KVCACHE_SPACE: "6"
+  VLLM_CPU_NUM_OF_RESERVED_CPU: "0"
+  VLLM_CPU_SGL_KERNEL: "1"
+  VLLM_ENGINE_ITERATION_TIMEOUT_S: "120"
+  VLLM_RPC_TIMEOUT: "100000"
+---
+# Source: vllm/templates/pvc.yaml
+# Copyright (C) 2025-2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: vllm-qwen-3-4b-cpu-pvc
+  namespace: default
+spec:
+  accessModes:
+  - ReadWriteOnce
+  resources:
+    requests:
+      storage: 250Gi # Adjust this according to model storage size
+---
+# Source: vllm/templates/service.yaml
+# Copyright (C) 2025-2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+apiVersion: v1
+kind: Service
+metadata:
+  name: vllm-qwen-3-4b-cpu-service
+  labels:
+    helm.sh/chart: vllm-0-latest
+    app.kubernetes.io/name: vllm
+    app.kubernetes.io/instance: vllm-qwen-3-4b-cpu
+    app.kubernetes.io/version: "0.5"
+    app.kubernetes.io/managed-by: Helm
+spec:
+  type: ClusterIP
+  ports:
+    - port: 80
+      targetPort: 2080
+      protocol: TCP
+      name: vllm
+  selector:
+    app.kubernetes.io/name: vllm
+    app.kubernetes.io/instance: vllm-qwen-3-4b-cpu
+---
+# Source: vllm/templates/deployment.yaml
+# Copyright (C) 2025-2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: vllm-qwen-3-4b-cpu
+  labels:
+    helm.sh/chart: vllm-0-latest
+    app.kubernetes.io/name: vllm
+    app.kubernetes.io/instance: vllm-qwen-3-4b-cpu
+    app.kubernetes.io/version: "0.5"
+    app.kubernetes.io/managed-by: Helm
+spec:
+  selector:
+    matchLabels:
+      app.kubernetes.io/name: vllm
+      app.kubernetes.io/instance: vllm-qwen-3-4b-cpu
+  template:
+    metadata:
+      annotations:
+        balloon.balloons.resource-policy.nri.io: "performance"
+      labels:
+        app.kubernetes.io/name: vllm
+        app.kubernetes.io/instance: vllm-qwen-3-4b-cpu
+        name: vllm
+    spec:
+      #
+      # serviceAccountName: default
+      securityContext:
+        fsGroup: 1001
+      containers:
+        - name: vllm
+          envFrom:
+            - configMapRef:
+                name: vllm-qwen-3-4b-cpu-config
+            - secretRef:
+                name: vllm-qwen-3-4b-cpu-secret
+          securityContext:
+            allowPrivilegeEscalation: false
+            capabilities:
+              add:
+              - SYS_NICE
+              drop:
+              - ALL
+            readOnlyRootFilesystem: true
+            runAsNonRoot: true
+            runAsUser: 1001
+            seccompProfile:
+              type: RuntimeDefault
+          image: "vllm/vllm-openai-cpu:v0.28.0@sha256:197354e6475e638eb5ccca09beb078488046a82de0e204e7429bf1a804df9647"
+          imagePullPolicy: IfNotPresent
+          # command:
+          #   - /bin/bash
+          #   - -c
+          #   - |
+          #     python3 -m vllm.entrypoints.openai.api_server --dtype bfloat16  --model Qwen/Qwen3-4B --port 2080 --tensor-parallel-size 1 --block-size 128 --max-model-len 8192 --disable-log-requests
+          args:
+            - "--block-size"
+            - "128"
+            - "--dtype"
+            - "bfloat16"
+            - "--distributed_executor_backend"
+            - "mp"
+            - "--enable_chunked_prefill"
+            - "--enforce-eager"
+            - "--max-num-batched-tokens"
+            - "2048"
+            - "--max-num-seqs"
+            - "256"
+            - "--enable-prefix-caching"
+            # CPU deployment - add parallelism settings
+            - "--tensor-parallel-size"
+            - "1"
+            - "--pipeline-parallel-size"
+            - "1"
+            - "--model"
+            - "Qwen/Qwen3-4B"
+            - --served-model-name
+            - "Qwen/Qwen3-4B"
+            - "--port"
+            - "2080"
+          volumeMounts:
+            - mountPath: /data
+              name: model-volume
+            - mountPath: /dev/shm
+              name: shm
+            - mountPath: /tmp
+              name: tmp
+            - mountPath: /root/.cache/vllm
+              name: vllm-cache
+          ports:
+            - name: http
+              containerPort: 2080
+              protocol: TCP
+          readinessProbe:
+            failureThreshold: 98
+            httpGet:
+              path: /health
+              port: http
+            initialDelaySeconds: 120
+            periodSeconds: 30
+          resources:
+            requests:
+              cpu: "16"
+              memory: "22Gi"
+            limits:
+              cpu: "16"
+              memory: "22Gi"
+
+      volumes:
+        - name: model-volume
+          persistentVolumeClaim:
+            claimName: vllm-qwen-3-4b-cpu-pvc
+        - name: shm
+          emptyDir:
+            medium: Memory
+            sizeLimit: 1Gi
+        - name: tmp
+          emptyDir: {}
+        - name: vllm-cache
+          emptyDir: {}
+      affinity:
+        nodeAffinity:
+          requiredDuringSchedulingIgnoredDuringExecution:
+            nodeSelectorTerms:
+            - matchExpressions:
+              - key: ei-inference-eligible
+                operator: In
+                values:
+                - "true"
+      # extra time to finish processing buffered requests on CPU before pod is forcibly terminated
+      terminationGracePeriodSeconds: 120
+---
+# Source: vllm/templates/apisixroutes.yaml
+# Copyright (C) 2025-2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+---
+# Source: vllm/templates/chat-template-configmap.yaml
+# Copyright (C) 2025-2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+---
+# Source: vllm/templates/ingress.yaml
+# Copyright (C) 2025-2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+# Please edit the object below. Lines beginning with a '#' will be ignored,
+# and an empty file will abort the edit. If an error occurs while saving this file will be
+# reopened with the relevant failures.
+---
+# Source: vllm/templates/ingress_eks.yaml
+# Copyright (C) 2025-2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+---
+# Source: vllm/templates/route.yaml
+# Copyright (C) 2025-2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+---
+# Source: vllm/templates/servicemonitor.yaml
+# Copyright (C) 2025-2026 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0