[UPDATE][fishspeechv3][1.1.9]

fishspeechv3/Chart.yaml

@@ -3,4 +3,4 @@ appVersion: 'webui-cuda-cu129-20260422'
 description: Fish Speech - Open-Source Multilingual Text-to-Speech with Voice Cloning
 name: fishspeechv3
 type: application
-version: 1.1.8
+version: '1.1.9'

fishspeechv3/OlaresManifest.yaml

@@ -7,7 +7,7 @@ metadata:
   description: SOTA Open Source TTS
   appid: fishspeechv3
   title: Fish Speech
-  version: '1.1.8'
+  version: '1.1.9'
   categories:
   - AI
 
@@ -74,12 +74,20 @@ spec:
     - GUI Inference: Offers a PyQt6 graphical interface that works seamlessly with the API server. Supports Linux, Windows, and macOS.
 
   upgradeDescription: |
-    Switched model provisioning to the shared llm-init downloader.
+    Storage: durable volumes moved from `appCache` to `appData`. On first start after upgrade, initContainers copy legacy appCache contents into appData when the destination is empty (marker: `.migration_from_appcache_done`). Old appCache data is left intact for rollback.
+
+    Drop legacy `sysVersion >= 1.12.3` appCache path branching; Olares minimum is now 1.12.6.
+
+    `HF_TOKEN` / `HF_ENDPOINT` now map from `OLARES_SYSTEM_HUGGINGFACE_TOKEN` / `OLARES_SYSTEM_HUGGINGFACE_SERVICE` (system-level), not user-level Hugging Face env.
+
+    Bump downloader image to `beclab/llm-init:v1.3.1`.
+
+    User data root moved from appCache to appData; compile-cache stays on appCache.
 
     **What's Changed**
 
     - **Download architecture**: replaced the bespoke `harveyff-hf-downloader`
-      Job with an `llm-init:v1.2.4` download-only Deployment
+      Job with an `llm-init:v1.3.1` download-only Deployment
       (`fishspeechv3download`). The fishaudio/s2-pro checkpoint now lands in the
       shared cross-app Hugging Face cache (App Common, `/cache/hf/hub`) instead
       of a per-app App Data copy, so it is de-duplicated across apps and survives
@@ -123,15 +131,15 @@ spec:
   limitedDisk: 50Gi
 envs:
 - envName: HF_TOKEN
-  required: true
+  required: false
   applyOnChange: true
   valueFrom:
-    envName: OLARES_USER_HUGGINGFACE_TOKEN
+    envName: OLARES_SYSTEM_HUGGINGFACE_TOKEN
 - envName: HF_ENDPOINT
-  required: true
+  required: false
   applyOnChange: true
   valueFrom:
-    envName: OLARES_USER_HUGGINGFACE_SERVICE
+    envName: OLARES_SYSTEM_HUGGINGFACE_SERVICE
 
 options:
   shared: true

fishspeechv3/i18n/en-US/OlaresManifest.yaml

@@ -31,4 +31,12 @@ spec:
     - GUI Inference: Offers a PyQt6 graphical interface that works seamlessly with the API server. Supports Linux, Windows, and macOS.
 
   upgradeDescription: |
-    Switched model provisioning to the shared llm-init downloader: replaced the bespoke downloader Job with an `llm-init:v1.2.4` download-only Deployment (`fishspeechv3download`) that fetches fishaudio/s2-pro into the shared cross-app Hugging Face cache (App Common, `/cache/hf/hub`); the WebUI now waits on the downloader's `/readyz` and symlinks `checkpoints/s2-pro` to the HF snapshot. The `fishspeechclient` nginx entrance serves the download dashboard and auto-reloads into the Gradio WebUI when ready. Added the `appCommon` permission. Previously migrated Fish Speech to the apiVersion v3 shared app pattern (fishspeechv2 → fishspeechv3).
+    Storage: durable volumes moved from `appCache` to `appData`. On first start after upgrade, initContainers copy legacy appCache contents into appData when the destination is empty (marker: `.migration_from_appcache_done`). Old appCache data is left intact for rollback.
+
+    Drop legacy `sysVersion >= 1.12.3` appCache path branching; Olares minimum is now 1.12.6.
+
+    `HF_TOKEN` / `HF_ENDPOINT` now map from `OLARES_SYSTEM_HUGGINGFACE_TOKEN` / `OLARES_SYSTEM_HUGGINGFACE_SERVICE` (system-level), not user-level Hugging Face env.
+
+    Bump downloader image to `beclab/llm-init:v1.3.1`.
+
+    User data root moved from appCache to appData; compile-cache stays on appCache.

fishspeechv3/i18n/zh-CN/OlaresManifest.yaml

@@ -26,4 +26,12 @@ spec:
     - 部署友好： 轻松设置推理服务器，原生支持 Linux、Windows（MacOS 即将推出），最小化速度损失。
 
   upgradeDescription: |
-    将模型下载切换为共享的 llm-init 下载器：以 `llm-init:v1.2.4` 下载专用 Deployment（`fishspeechv3download`）替换原有的下载器 Job，将 fishaudio/s2-pro 下载到跨应用共享的 Hugging Face 缓存（App Common，`/cache/hf/hub`）；WebUI 改为等待下载器的 `/readyz`，并将 `checkpoints/s2-pro` 软链到 HF 快照。`fishspeechclient` nginx 入口在下载期间展示下载进度页，就绪后自动跳转到 Gradio WebUI。新增 `appCommon` 权限。此前已将 Fish Speech 迁移到 apiVersion v3 共享应用模式（fishspeechv2 → fishspeechv3）。
+    存储：持久卷从 `appCache` 迁移到 `appData`。升级后首次启动时，若 appData 为空，initContainer 会将旧 appCache 内容复制到 appData（标记文件：`.migration_from_appcache_done`）；旧 appCache 数据保留以便回滚。
+
+    移除针对 Olares < 1.12.3 的 appCache 路径分支逻辑；当前最低支持版本为 1.12.6。
+
+    `HF_TOKEN` / `HF_ENDPOINT` 改为从系统级 `OLARES_SYSTEM_HUGGINGFACE_TOKEN` / `OLARES_SYSTEM_HUGGINGFACE_SERVICE` 获取，不再使用用户级 Hugging Face 环境变量。
+
+    下载器镜像升级为 `beclab/llm-init:v1.3.1`。
+
+    用户数据根目录从 appCache 迁移到 appData；compile-cache 仍保留在 appCache。

fishspeechv3/templates/server.yaml

@@ -24,31 +24,62 @@ spec:
         io.kompose.service: fishspeech
     spec:
       initContainers:
-      - name: init-chmod-data
-        image: 'docker.io/beclab/aboveos-busybox:1.37.0'
-        command:
-          - sh
-          - '-c'
-          - |
-            chown -R 1000:1000 /app-checkpoints /app-references /app-data-root /tmp-gradio /app-cache
-            chmod -R 755 /app-checkpoints /app-references /app-data-root /tmp-gradio /app-cache
-        resources: {}
-        volumeMounts:
-          - name: checkpoints
-            mountPath: /app-checkpoints
-          - name: references
-            mountPath: /app-references
-          - name: data
-            mountPath: /app-data-root
-          - name: output
-            mountPath: /tmp-gradio
-          - name: app-cache
-            mountPath: /app-cache
-        terminationMessagePath: /dev/termination-log
-        terminationMessagePolicy: File
-        imagePullPolicy: IfNotPresent
-        securityContext:
-          runAsUser: 0
+        - name: migrate-appcache-to-appdata
+          image: "docker.io/beclab/busybox:1.37"
+          imagePullPolicy: IfNotPresent
+          command: ["sh", "-c"]
+          args:
+            - |
+              set -e
+              SOURCE=/old-data
+              DEST=/new-data
+              MARKER="${DEST}/.migration_from_appcache_done"
+              if [ -f "$MARKER" ]; then echo "[migrate] marker present, skip."; exit 0; fi
+              if [ -n "$(ls -A "$DEST" 2>/dev/null | grep -v '^\.migration_from_appcache_done$' || true)" ]; then
+                echo "[migrate] appData already has data, skip."
+                echo "skipped existing dest at $(date -Iseconds 2>/dev/null || date)" > "$MARKER"
+                exit 0
+              fi
+              if [ -n "$(ls -A "$SOURCE" 2>/dev/null || true)" ]; then
+                echo "[migrate] copying appCache -> appData..."
+                cp -a "$SOURCE"/. "$DEST"/
+                echo "migrated from appCache at $(date -Iseconds 2>/dev/null || date)" > "$MARKER"
+                echo "[migrate] done; legacy appCache left intact for rollback."
+              else
+                echo "[migrate] no legacy data, fresh install."
+                echo "fresh install at $(date -Iseconds 2>/dev/null || date)" > "$MARKER"
+              fi
+          volumeMounts:
+            - name: data-legacy
+              mountPath: /old-data
+              readOnly: true
+            - name: data
+              mountPath: /new-data
+        - name: init-chmod-data
+          image: 'docker.io/beclab/aboveos-busybox:1.37.0'
+          command:
+            - sh
+            - '-c'
+            - |
+              chown -R 1000:1000 /app-checkpoints /app-references /app-data-root /tmp-gradio /app-cache
+              chmod -R 755 /app-checkpoints /app-references /app-data-root /tmp-gradio /app-cache
+          resources: {}
+          volumeMounts:
+            - name: checkpoints
+              mountPath: /app-checkpoints
+            - name: references
+              mountPath: /app-references
+            - name: data
+              mountPath: /app-data-root
+            - name: output
+              mountPath: /tmp-gradio
+            - name: app-cache
+              mountPath: /app-cache
+          terminationMessagePath: /dev/termination-log
+          terminationMessagePolicy: File
+          imagePullPolicy: IfNotPresent
+          securityContext:
+            runAsUser: 0
       containers:
       - name: fishspeech
         image: "docker.io/beclab/harveyff-fishaudio-fish-speech:webui-cuda-cu129-20260422"
@@ -184,6 +215,10 @@ spec:
       - hostPath:
           path: '{{ .Values.userspace.appCache }}/data'
           type: DirectoryOrCreate
+        name: data-legacy
+      - hostPath:
+          path: '{{ .Values.userspace.appData }}/data'
+          type: DirectoryOrCreate
         name: data
       # checkpoints/s2-pro is now an ephemeral symlink into the shared HF cache,
       # rebuilt on every start; the real model bytes live in hf-cache below.
@@ -269,17 +304,16 @@ spec:
             runAsUser: 0
       containers:
         - name: llm-init
-          # v1.2.4: nullengine download-only, default LFS (hf_xet off unless
+          # v1.3.1: nullengine download-only, default LFS (hf_xet off unless
           # HF_ENABLE_XET=true) for stable download memory.
-          image: docker.io/beclab/llm-init:v1.2.4
+          image: docker.io/beclab/llm-init:v1.3.1
           imagePullPolicy: IfNotPresent
           env:
             # Empty ENGINE_KIND selects download-only mode (nullengine).
             - name: ENGINE_KIND
               value: ""
-            # TEMPORARY (llm-init v1.2.4 image): download-only still requires
-            # MODEL_NAME + MODEL_MODE placeholders; the repo to fetch comes
-            # entirely from MODEL_SOURCE. Drop both once the make-optional patch ships.
+            # Download-only still requires MODEL_NAME + MODEL_MODE placeholders;
+            # the repo to fetch comes entirely from MODEL_SOURCE.
             - name: MODEL_NAME
               value: "fishaudio/s2-pro"
             - name: MODEL_MODE
@@ -294,7 +328,7 @@ spec:
             - name: HF_TOKEN
               value: "{{ $hfToken }}"
 {{- end }}
-            # Keep concurrency low; LFS is the image default in v1.2.4.
+            # Keep concurrency low; LFS is the image default.
             - name: MAX_CONCURRENT_DOWNLOADS
               value: "1"
             - name: RUN_DIR