move kokoro to separate deployment

apps/llama/configs/config.yaml

@@ -20,7 +20,6 @@ hooks:
     preload:
       - "Qwen3.5-0.8B-GGUF-nothink:Q4_K_XL"
       - "whisper-small"
-      - "outetts-tts"
 
 # matrix replaces groups (they are mutually exclusive).
 # The small 0.8B model runs alongside any LLM.
@@ -29,7 +28,6 @@ matrix:
   vars:
     q8: "Qwen3.5-0.8B-GGUF-nothink:Q4_K_XL"
     stt: "whisper-small"
-    tts: "outetts-tts"
     flux: "flux2-klein-4b:Q4_K_M"
     coder: "Qwen3-Coder-Next-GGUF:Q4_K_M"
     q35t: "Qwen3.5-35B-A3B-GGUF:Q4_K_M"
@@ -56,9 +54,9 @@ matrix:
 
   sets:
     # any LLM can run alongside the small always-on model + STT + TTS (all CPU, no VRAM cost)
-    with_q8: "(coder | q35t | q35nt | q35ht | q35hnt | q4t | q4nt | q4ht | q4hnt | g26xl | g26q2 | ge4xl | ge2xl | q36t | q36nt | haut | haunt | mtpt | mtpnt) & q8 & stt & tts"
-    # FLUX runs alone — evicts everything including q8, but keeps STT+TTS for voice during image gen
-    image_gen: "flux & stt & tts"
+    with_q8: "(coder | q35t | q35nt | q35ht | q35hnt | q4t | q4nt | q4ht | q4hnt | g26xl | g26q2 | ge4xl | ge2xl | q36t | q36nt | haut | haunt | mtpt | mtpnt) & q8 & stt"
+    # FLUX runs alone — evicts everything including q8, but keeps STT for voice during image gen
+    image_gen: "flux & stt"
 
 models:
   "Qwen3-Coder-Next-GGUF:Q4_K_M":
@@ -252,19 +250,6 @@ models:
         --threads 6
         --no-gpu
 
-  # TTS via OuteTTS 0.3 1B + WavTokenizer vocoder (CPU-only, always loaded)
-  # Models auto-downloaded from HuggingFace on first start
-  # OuteTTS 0.3 1B: ~1GB RAM, WavTokenizer: ~600MB RAM
-  # Exposes /v1/audio/speech compatible with OpenAI TTS API
-  "outetts-tts":
-    checkEndpoint: none
-    cmd: |
-      llama-server
-        -hf OuteAI/OuteTTS-0.3-1B-GGUF
-        -hff OuteTTS-0.3-1B-Q8_0.gguf
-        -mv /root/.cache/huggingface/hub/models--ggml-org--WavTokenizer/snapshots/0c97fdc098158ec9bf4e703cd5f81a5aa20520e6/WavTokenizer-Large-75-F16.gguf
-        -c 4096
-        ${cpu_args}
 
   # Image generation via stable-diffusion.cpp (sd-server)
   # Models must be pre-downloaded to /root/.cache/sd/

apps/llama/kokoro.yaml

@@ -0,0 +1,65 @@
+---
+apiVersion: apps/v1
+kind: Deployment
+metadata:
+  name: kokoro
+  namespace: llama
+spec:
+  replicas: 1
+  strategy:
+    type: Recreate
+  selector:
+    matchLabels:
+      app: kokoro
+  template:
+    metadata:
+      labels:
+        app: kokoro
+    spec:
+      # Prevent Kubernetes from auto-injecting KOKORO_PORT=tcp://... (legacy
+      # Docker-link-style env vars from same-namespace Services), which Kokoro
+      # tries to parse as an integer and fails.
+      enableServiceLinks: false
+      containers:
+        - name: kokoro
+          # OpenAI-compatible Kokoro TTS server, CPU-only PyTorch backend
+          # Exposes POST /v1/audio/speech with multiple voices and streaming
+          image: hwdsl2/kokoro-server:latest@sha256:42886b8720e901f7e31aba2050cd03867767eb9f609bbc38fe93852e72f0feeb
+          ports:
+            - containerPort: 8880
+              name: http
+              protocol: TCP
+          env:
+            # Default voice (can be overridden per-request)
+            - name: KOKORO_VOICE
+              value: "af_heart"
+          volumeMounts:
+            # Persistent cache for the ~320MB Kokoro model
+            - name: cache
+              mountPath: /var/lib/kokoro
+          resources:
+            requests:
+              memory: "2Gi"
+              cpu: "500m"
+            limits:
+              memory: "6Gi"
+      volumes:
+        - name: cache
+          persistentVolumeClaim:
+            claimName: kokoro-cache-lvmssd
+---
+apiVersion: v1
+kind: Service
+metadata:
+  name: kokoro
+  namespace: llama
+spec:
+  type: ClusterIP
+  ports:
+    - name: http
+      port: 8880
+      targetPort: 8880
+      protocol: TCP
+  selector:
+    app: kokoro
+

apps/llama/kustomization.yaml

@@ -6,7 +6,9 @@ resources:
   - auth-proxy.yaml
   - ingress.yaml
   - pvc-ssd.yaml
+  - pvc-kokoro-ssd.yaml
   - deployment.yaml
+  - kokoro.yaml
 configMapGenerator:
   - name: llama-swap
     namespace: llama

apps/llama/pvc-kokoro-ssd.yaml

@@ -0,0 +1,46 @@
+---
+apiVersion: local.openebs.io/v1alpha1
+kind: LVMVolume
+metadata:
+  labels:
+    kubernetes.io/nodename: anapistula-delrosalae
+  name: kokoro-cache-lvmssd
+  namespace: openebs
+spec:
+  capacity: "2147483648"
+  ownerNodeID: anapistula-delrosalae
+  shared: "yes"
+  thinProvision: "no"
+  vgPattern: ^openebs-ssd$
+  volGroup: openebs-ssd
+---
+kind: PersistentVolume
+apiVersion: v1
+metadata:
+  name: kokoro-cache-lvmssd
+spec:
+  capacity:
+    storage: 2Gi
+  accessModes:
+    - ReadWriteOnce
+  persistentVolumeReclaimPolicy: Retain
+  storageClassName: ssd-lvmpv
+  volumeMode: Filesystem
+  csi:
+    driver: local.csi.openebs.io
+    fsType: btrfs
+    volumeHandle: kokoro-cache-lvmssd
+---
+kind: PersistentVolumeClaim
+apiVersion: v1
+metadata:
+  name: kokoro-cache-lvmssd
+  namespace: llama
+spec:
+  accessModes:
+    - ReadWriteOnce
+  resources:
+    requests:
+      storage: 2Gi
+  storageClassName: ssd-lvmpv
+  volumeName: kokoro-cache-lvmssd

apps/openwebui/release.yaml

@@ -86,16 +86,17 @@ spec:
       - name: AUDIO_STT_SUPPORTED_CONTENT_TYPES
         value: "audio/wav,audio/wave"
       # TTS via OuteTTS (routed through llama-swap)
+      # TTS via dedicated Kokoro server (CPU-only, separate pod)
       - name: AUDIO_TTS_ENGINE
         value: "openai"
       - name: AUDIO_TTS_OPENAI_API_BASE_URL
-        value: "http://llama.llama.svc.cluster.local:11434/v1"
+        value: "http://kokoro.llama.svc.cluster.local:8880/v1"
       - name: AUDIO_TTS_OPENAI_API_KEY
         value: "ignored"
       - name: AUDIO_TTS_MODEL
-        value: "outetts-tts"
+        value: "kokoro"
       - name: AUDIO_TTS_VOICE
-        value: "default"
+        value: "af_heart"
       # Image generation via llama-swap sd-server
       - name: ENABLE_IMAGE_GENERATION
         value: "true"