feat(llama-server): split embeddings off, move chat to Qwen3.6-35B-A3B

Two defects surfaced once the endpoint was live, and both trace back to
router mode.

The embedding model was selectable in the chat UI. Every model on the port
appears in the picker, and choosing the embedding one returns fluent nonsense
("capital capital capital ...") rather than an error, because --embeddings does
not actually refuse chat completions. llama.cpp offers no way to hide a model
from the list: --tags is informational, --ui-config only sets UI preferences.
So the embedding model moves to its own Deployment on a ClusterIP-only Service
with --no-webui, which is what keeps it out of the picker.

Readiness was also wrong. The router answers /health 200 as soon as it listens,
while models still download behind it, so the pod went Ready seconds after start
and traffic arrived at a server that could not answer — that is how the
embedding model got reached at all. A single-model server returns 503 until the
weights are resident, which is what the probes needed all along.

Splitting fixes both, at the cost of router mode's single endpoint. Consumers
now use llama-server for chat and llama-embed for embeddings; both share one
PVC as LLAMA_CACHE, so nothing is downloaded twice.

The chat model changes to a mixture-of-experts. Llama-3.1-8B was picked only
because it was the model benchmarked, which is continuity rather than a quality
argument, and it showed: fluent but wrong. Since decode is set by bytes read per
token, an MoE reading ~1.85 GB/token beats a dense model reading ~16.5 GB.
Measured on k8s-1 with this image:

  tg128     11.59 tok/s   (dense 8B: 10.78)
  pp512    120.93 t/s
  pp2048   126.47 t/s
  pp8192    92.79 t/s
  pp16384   69.84 t/s

Four times the model, slightly faster. The prefill falloff is sublinear, so the
Gated Delta Net collapse of llama.cpp#27638 does not occur here and long
documents stay viable; context therefore goes to 16k, which is cheap because
only 10 of 40 layers use full attention.

GLM-5.2/5.3 were evaluated and are impossible on this hardware at 754B
parameters (~400 GB), as is GLM-5.3-Flash at 321B with an architecture the
pinned build does not implement. Dense Qwen3.8-27B reads 16.5 GB/token and
manages 3.3 tok/s.

Memory goes to 24Gi/32Gi: Vulkan allocates the iGPU working set from system RAM
and it counts against the pod cgroup, measured at 7.54 GiB RSS for 5.2 GiB of
weights. --ubatch-size 1024 because 2048 and above fail with ErrorDeviceLost on
this Vulkan/Mesa stack.

Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>

kubernetes/RUNBOOK-cold-rebuild.md

@@ -119,9 +119,11 @@ against do not exist, and the node ends up with no address on any interface.
 Beyond that it carries:
 
 - `siderolabs/intel-ucode` — no hypervisor supplies microcode here
-- `siderolabs/i915` — the Alder Lake-P Iris Xe. Nothing requests `/dev/dri`
-  today; present so wanting hardware transcoding later is a config change
-  rather than another rolling reinstall
+- `siderolabs/i915` — the Alder Lake-P Iris Xe. The Intel GPU device plugin
+  (`infrastructure/intel-gpu-plugin`) turns this into the schedulable resource
+  `gpu.intel.com/i915`, which `apps/llama-server` requests for LLM inference
+  and which any hardware-transcoding workload would use too. Without the
+  extension the plugin finds no `/dev/dri` and advertises nothing
 - `siderolabs/thunderbolt` — inert until `machine.kernel.modules` requests
   `thunderbolt`/`thunderbolt_net`, present for the same reason. Note its udev
   rule auto-authorises any Thunderbolt device attached once those modules load

kubernetes/apps/llama-server/deployment.yaml

@@ -0,0 +1,164 @@
+---
+# The chat endpoint, and the only one behind the ingress.
+#
+# Single-model rather than router mode. Router mode served chat and embeddings
+# from one port, but two things made that wrong here:
+#
+#   1. Its /health answers 200 as soon as the router process listens, while the
+#      models are still downloading behind it. Kubernetes marked the pod Ready
+#      within seconds of start, so traffic arrived at a server that could not
+#      answer. In single-model mode /health returns 503 until the model is
+#      resident, which is what the probes below actually need.
+#   2. Every model on the port is selectable in the built-in WebUI, including
+#      the embedding model. Asking an embedding model to chat returns fluent
+#      nonsense ("capital capital capital ...") rather than an error, and
+#      `--embeddings` does not prevent it. There is no flag to hide a model
+#      from the list — `--tags` is informational and `--ui-config` only sets UI
+#      preferences — so the only way to keep it out of the picker is to keep it
+#      off this endpoint. It runs in embed.yaml instead.
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: llama-server
+  namespace: ai
+  labels:
+    app.kubernetes.io/name: llama-server
+spec:
+  replicas: 1
+  # A rolling update would load a second copy of the weights onto the same
+  # iGPU and into the same node's RAM before releasing the first.
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      app.kubernetes.io/name: llama-server
+  template:
+    metadata:
+      labels:
+        app.kubernetes.io/name: llama-server
+    spec:
+      securityContext:
+        runAsNonRoot: true
+        runAsUser: 1000
+        runAsGroup: 1000
+        # Makes the CephFS volume group-writable so the weights can be cached.
+        fsGroup: 1000
+        seccompProfile:
+          type: RuntimeDefault
+      containers:
+        - name: llama-server
+          image: ghcr.io/ggml-org/llama.cpp:server-vulkan-b10731
+          args:
+            # The one place this model is named; llama-server resolves the repo
+            # and downloads it itself, so there is nothing to keep in sync.
+            #
+            # A 35B mixture-of-experts with only ~3B parameters active per
+            # token. Decode on this hardware is bound by how many bytes are read
+            # per token rather than by parameter count, and this model reads
+            # ~1.85 GB/token where a dense 27B reads ~16.5 GB — so it is both
+            # far larger and faster than any dense model that would fit.
+            #
+            # Measured here (llama-bench, k8s-1, this image): tg128 11.59 tok/s,
+            # against 10.78 for a dense 8B. Prefill 120.9 t/s at 512 tokens and
+            # 69.8 at 16384 — a sublinear falloff, so the Gated Delta Net
+            # prefill collapse of llama.cpp#27638 does not occur on this stack.
+            - --hf-repo
+            - ggml-org/Qwen3.6-35B-A3B-GGUF:Q4_K_M
+            # Every layer on the Intel iGPU. Full offload matches CPU throughput
+            # while leaving the node's 14 cores free for the rest of the
+            # cluster; the iGPU is here for that isolation, not for a speed-up.
+            - --n-gpu-layers
+            - "99"
+            # Decode runs on the GPU, so extra threads buy nothing. At 11.59
+            # tok/s this model moves 21.4 GB/s against the 53.9 GB/s a dense
+            # model sustains here, so it is the iGPU's 96 EUs working through
+            # many small per-expert matmuls that sets the pace, not bandwidth.
+            - --threads
+            - "4"
+            # 16k rather than 8k because the measured prefill falloff is mild
+            # and this model's KV cache is unusually cheap — only 10 of its 40
+            # layers use full attention.
+            - --ctx-size
+            - "16384"
+            # A micro-batch of 2048 or more makes this Vulkan/Mesa stack fail
+            # with ErrorDeviceLost.
+            - --ubatch-size
+            - "1024"
+            # The server otherwise allows every CORS origin and warns about it
+            # at startup. Without this, any page a LAN browser visits could call
+            # this unauthenticated API from the user's machine. Consumers talk
+            # to it server-side; the WebUI is same-origin and unaffected.
+            - --cors-origins
+            - localhost
+          env:
+            # Where --hf-repo caches the weights. Pointing it at the PVC makes
+            # the download a one-time cost, shared with llama-embed, and keeps
+            # the server off a home directory it cannot write as UID 1000.
+            - name: LLAMA_CACHE
+              value: /models
+          ports:
+            - name: http
+              containerPort: 8080
+          securityContext:
+            allowPrivilegeEscalation: false
+            capabilities:
+              drop:
+                - ALL
+          # Vulkan allocates the iGPU's working set from system RAM, and that
+          # counts against the pod's cgroup: a 5.2 GiB model measured 7.54 GiB
+          # RSS. Sized as the 19 GiB of weights plus that overhead, plus the KV
+          # cache — which is unusually cheap here, since only 10 of the 40
+          # layers use full attention.
+          resources:
+            requests:
+              cpu: 500m
+              memory: 24Gi
+              gpu.intel.com/i915: "1"
+            limits:
+              cpu: "6"
+              memory: 32Gi
+              gpu.intel.com/i915: "1"
+          # /health is 503 while the model loads and 200 once it is resident, so
+          # the pod only goes Ready when it can actually answer. On a cold PVC
+          # that includes pulling ~19 GiB from Hugging Face, hence the long
+          # startup budget (40 min); liveness takes over only after that.
+          startupProbe:
+            httpGet:
+              path: /health
+              port: http
+            periodSeconds: 10
+            failureThreshold: 240
+          readinessProbe:
+            httpGet:
+              path: /health
+              port: http
+            periodSeconds: 10
+          livenessProbe:
+            httpGet:
+              path: /health
+              port: http
+            periodSeconds: 30
+            timeoutSeconds: 10
+            failureThreshold: 6
+          volumeMounts:
+            - name: models
+              mountPath: /models
+      volumes:
+        - name: models
+          persistentVolumeClaim:
+            claimName: llama-server-models
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: llama-server
+  namespace: ai
+  labels:
+    app.kubernetes.io/name: llama-server
+spec:
+  selector:
+    app.kubernetes.io/name: llama-server
+  ports:
+    - name: http
+      port: 8080
+      targetPort: http

kubernetes/apps/llama-server/embed.yaml

@@ -0,0 +1,119 @@
+---
+# The embedding endpoint, for RAG and semantic search.
+#
+# Deliberately separate from llama-server and deliberately not behind the
+# ingress: an embedding model has no generative training, and when it shares a
+# port with the chat model the built-in WebUI lists it as a choice. Selecting it
+# returns fluent nonsense rather than an error, because `--embeddings` does not
+# actually refuse chat completions. Keeping it on its own ClusterIP-only Service
+# is what makes it unselectable, since llama.cpp has no way to hide a model from
+# the picker.
+#
+# Reachable in-cluster at http://llama-embed.ai.svc.cluster.local:8080/v1/embeddings.
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: llama-embed
+  namespace: ai
+  labels:
+    app.kubernetes.io/name: llama-embed
+spec:
+  replicas: 1
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      app.kubernetes.io/name: llama-embed
+  template:
+    metadata:
+      labels:
+        app.kubernetes.io/name: llama-embed
+    spec:
+      securityContext:
+        runAsNonRoot: true
+        runAsUser: 1000
+        runAsGroup: 1000
+        fsGroup: 1000
+        seccompProfile:
+          type: RuntimeDefault
+      containers:
+        - name: llama-embed
+          image: ghcr.io/ggml-org/llama.cpp:server-vulkan-b10731
+          args:
+            - --hf-repo
+            - Qwen/Qwen3-Embedding-0.6B-GGUF:Q8_0
+            # Serves POST /v1/embeddings.
+            - --embeddings
+            - --n-gpu-layers
+            - "99"
+            - --threads
+            - "2"
+            - --ctx-size
+            - "2048"
+            # No browser ever talks to this Service, so the WebUI is dead weight
+            # and one more unauthenticated surface.
+            - --no-webui
+          env:
+            # Same cache as llama-server, so the two share one copy of anything
+            # they have in common and neither re-downloads the other's weights.
+            - name: LLAMA_CACHE
+              value: /models
+          ports:
+            - name: http
+              containerPort: 8080
+          securityContext:
+            allowPrivilegeEscalation: false
+            capabilities:
+              drop:
+                - ALL
+          resources:
+            requests:
+              cpu: 250m
+              memory: 2Gi
+              gpu.intel.com/i915: "1"
+            limits:
+              cpu: "2"
+              memory: 4Gi
+              gpu.intel.com/i915: "1"
+          # ~600 MiB rather than ~5 GiB, so a much shorter budget than the chat
+          # model needs, but still sized for a cold download.
+          startupProbe:
+            httpGet:
+              path: /health
+              port: http
+            periodSeconds: 10
+            failureThreshold: 60
+          readinessProbe:
+            httpGet:
+              path: /health
+              port: http
+            periodSeconds: 10
+          livenessProbe:
+            httpGet:
+              path: /health
+              port: http
+            periodSeconds: 30
+            timeoutSeconds: 10
+            failureThreshold: 6
+          volumeMounts:
+            - name: models
+              mountPath: /models
+      volumes:
+        - name: models
+          persistentVolumeClaim:
+            claimName: llama-server-models
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: llama-embed
+  namespace: ai
+  labels:
+    app.kubernetes.io/name: llama-embed
+spec:
+  selector:
+    app.kubernetes.io/name: llama-embed
+  ports:
+    - name: http
+      port: 8080
+      targetPort: http

kubernetes/apps/llama-server/ingress.yaml

@@ -0,0 +1,28 @@
+---
+# LAN-only: no `dns.renner.uno/public` label, so the Cloudflare external-dns
+# instance skips it and only the UDM instance publishes the record. The API is
+# unauthenticated, so it must not be reachable from the internet.
+apiVersion: networking.k8s.io/v1
+kind: Ingress
+metadata:
+  name: llama-server
+  namespace: ai
+  annotations:
+    cert-manager.io/cluster-issuer: letsencrypt-cluster
+spec:
+  ingressClassName: traefik
+  tls:
+    - hosts:
+        - llm.renner.uno
+      secretName: llama-server-tls
+  rules:
+    - host: llm.renner.uno
+      http:
+        paths:
+          - path: /
+            pathType: Prefix
+            backend:
+              service:
+                name: llama-server
+                port:
+                  number: 8080

kubernetes/apps/llama-server/kustomization.yaml

@@ -0,0 +1,7 @@
+apiVersion: kustomize.config.k8s.io/v1beta1
+kind: Kustomization
+resources:
+  - pvc.yaml
+  - deployment.yaml
+  - embed.yaml
+  - ingress.yaml

kubernetes/apps/llama-server/pvc.yaml

@@ -0,0 +1,21 @@
+---
+# Holds the GGUF weights, downloaded once by the Deployment's initContainer.
+# Prune-disabled so retiring this Kustomization does not throw away several
+# GiB of model files that would have to be pulled from Hugging Face again.
+#
+# ReadWriteMany so a second replica or a one-off benchmark Job can mount the
+# same weights; the models are read-only at inference time.
+apiVersion: v1
+kind: PersistentVolumeClaim
+metadata:
+  name: llama-server-models
+  namespace: ai
+  annotations:
+    kustomize.toolkit.fluxcd.io/prune: disabled
+spec:
+  accessModes:
+    - ReadWriteMany
+  storageClassName: ceph-filesystem
+  resources:
+    requests:
+      storage: 40Gi

kubernetes/clusters/homelab/llama-server.yaml

@@ -0,0 +1,22 @@
+apiVersion: kustomize.toolkit.fluxcd.io/v1
+kind: Kustomization
+metadata:
+  name: llama-server
+  namespace: flux-system
+spec:
+  interval: 10m
+  retryInterval: 1m
+  # On a cold PVC the chat model pulls ~19 GiB of GGUF weights from Hugging
+  # Face and then loads onto the iGPU before /health goes green, so a first
+  # start is far slower than a normal app's.
+  timeout: 60m
+  sourceRef:
+    kind: GitRepository
+    name: flux-system
+  path: ./kubernetes/apps/llama-server
+  prune: true
+  wait: true
+  dependsOn:
+    - name: infrastructure
+    - name: infrastructure-crds
+    - name: infrastructure-ceph-cluster

kubernetes/infrastructure/namespaces/namespaces.yaml

@@ -183,3 +183,15 @@ apiVersion: v1
 kind: Namespace
 metadata:
   name: woodpecker
+---
+# Local LLM inference (kubernetes/apps/llama-server). Declared here, like
+# `woodpecker` above, so the `infrastructure` KS creates it before
+# `infrastructure-crds` applies the namespace's CiliumNetworkPolicies.
+#
+# No pod-security labels: the workload needs no hostPath or privilege — the
+# Intel GPU device plugin passes the iGPU in through the device-plugin API —
+# so the cluster's default `baseline` enforcement is enough.
+apiVersion: v1
+kind: Namespace
+metadata:
+  name: ai

kubernetes/infrastructure/network-policies/ai-ns.yaml

@@ -0,0 +1,96 @@
+apiVersion: cilium.io/v2
+kind: CiliumNetworkPolicy
+metadata:
+  name: default-deny-ingress
+  namespace: ai
+spec:
+  endpointSelector: {}
+  ingress:
+    - fromEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": ai
+    - fromEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": ingress
+            app.kubernetes.io/name: traefik
+    - fromEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": monitoring
+    - fromEntities:
+        - host
+        - remote-node
+    # ICMP — all standard ICMPv4 + ICMPv6 from any cluster pod. See
+    # home-assistant-ns.yaml for the rationale (CT-RELATED path not
+    # passing related ICMP empirically; per-namespace shape, mirroring
+    # commit 3c517a0's verified-safe pattern with the 66dbe9f type set).
+    - fromEndpoints:
+        - matchExpressions:
+            - key: io.kubernetes.pod.namespace
+              operator: Exists
+      icmps:
+        - fields:
+            - { type: 0,   family: IPv4 }
+            - { type: 3,   family: IPv4 }
+            - { type: 8,   family: IPv4 }
+            - { type: 11,  family: IPv4 }
+            - { type: 12,  family: IPv4 }
+            - { type: 1,   family: IPv6 }
+            - { type: 2,   family: IPv6 }
+            - { type: 3,   family: IPv6 }
+            - { type: 4,   family: IPv6 }
+            - { type: 128, family: IPv6 }
+            - { type: 129, family: IPv6 }
+            - { type: 133, family: IPv6 }
+            - { type: 134, family: IPv6 }
+            - { type: 135, family: IPv6 }
+            - { type: 136, family: IPv6 }
+            - { type: 143, family: IPv6 }
+  egress:
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": ai
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": ingress
+            app.kubernetes.io/name: traefik
+    # rook-ceph: the model weights live on a CephFS RWX volume
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": rook-ceph
+    - toEndpoints:
+        - matchLabels:
+            "k8s:io.kubernetes.pod.namespace": kube-system
+            k8s-app: kube-dns
+      toPorts:
+        - ports:
+            - port: "53"
+              protocol: UDP
+    # host/remote-node: kubelet probes + node-local reach (Ceph mons on host net)
+    # world: image pulls and llama-server's Hugging Face model download
+    - toEntities:
+        - host
+        - remote-node
+        - world
+    # ICMP — egress mirror; see ingress block above.
+    - toEndpoints:
+        - matchExpressions:
+            - key: io.kubernetes.pod.namespace
+              operator: Exists
+      icmps:
+        - fields:
+            - { type: 0,   family: IPv4 }
+            - { type: 3,   family: IPv4 }
+            - { type: 8,   family: IPv4 }
+            - { type: 11,  family: IPv4 }
+            - { type: 12,  family: IPv4 }
+            - { type: 1,   family: IPv6 }
+            - { type: 2,   family: IPv6 }
+            - { type: 3,   family: IPv6 }
+            - { type: 4,   family: IPv6 }
+            - { type: 128, family: IPv6 }
+            - { type: 129, family: IPv6 }
+            - { type: 133, family: IPv6 }
+            - { type: 134, family: IPv6 }
+            - { type: 135, family: IPv6 }
+            - { type: 136, family: IPv6 }
+            - { type: 143, family: IPv6 }

kubernetes/infrastructure/network-policies/kustomization.yaml

@@ -24,4 +24,5 @@ resources:
   - woodpecker-ns.yaml
   - trmnl-ns.yaml
   - wordpress-ns.yaml
+  - ai-ns.yaml
   - allow-cilium-health.yaml