From ab5cd96172a10be2e7bfeccec97ddb3677ad9e0e Mon Sep 17 00:00:00 2001
From: Harvey <harvey@bytetrade.io>
Date: Tue, 15 Sep 2026 12:06:55 +0800
Subject: [PATCH 1/3] Update llamacppllmbasev3 to tested 1.2.47 with CUDA
 b10969

---
 llamacppllmbasev3/Chart.yaml                  |  4 +-
 llamacppllmbasev3/OlaresManifest.yaml         | 82 +++++++++-------
 .../i18n/de-DE/OlaresManifest.yaml            | 40 ++++++++
 .../i18n/en-US/OlaresManifest.yaml            | 24 +----
 .../i18n/es-ES/OlaresManifest.yaml            | 40 ++++++++
 .../i18n/fr-FR/OlaresManifest.yaml            | 40 ++++++++
 .../i18n/it-IT/OlaresManifest.yaml            | 40 ++++++++
 .../i18n/ja-JP/OlaresManifest.yaml            | 40 ++++++++
 .../i18n/zh-CN/OlaresManifest.yaml            | 24 +----
 llamacppllmbasev3/templates/_helpers.tpl      | 50 ++++++++++
 llamacppllmbasev3/templates/llamacpp.yaml     | 97 ++++++++++++++++---
 llamacppllmbasev3/templates/llm-init.yaml     |  4 +-
 llamacppllmbasev3/values.yaml                 | 10 ++
 13 files changed, 402 insertions(+), 93 deletions(-)
 create mode 100644 llamacppllmbasev3/i18n/de-DE/OlaresManifest.yaml
 create mode 100644 llamacppllmbasev3/i18n/es-ES/OlaresManifest.yaml
 create mode 100644 llamacppllmbasev3/i18n/fr-FR/OlaresManifest.yaml
 create mode 100644 llamacppllmbasev3/i18n/it-IT/OlaresManifest.yaml
 create mode 100644 llamacppllmbasev3/i18n/ja-JP/OlaresManifest.yaml

diff --git a/llamacppllmbasev3/Chart.yaml b/llamacppllmbasev3/Chart.yaml
index a3f6cdcd1..f9d32c363 100644
--- a/llamacppllmbasev3/Chart.yaml
+++ b/llamacppllmbasev3/Chart.yaml
@@ -1,6 +1,6 @@
 apiVersion: v2
-appVersion: b10588
+appVersion: b10969
 description: Generic llama.cpp + llm-init base; the GGUF model is supplied at install time via env
 name: llamacppllmbasev3
 type: application
-version: 1.2.37
\ No newline at end of file
+version: 1.2.47
diff --git a/llamacppllmbasev3/OlaresManifest.yaml b/llamacppllmbasev3/OlaresManifest.yaml
index 01bf01017..3caa50aeb 100644
--- a/llamacppllmbasev3/OlaresManifest.yaml
+++ b/llamacppllmbasev3/OlaresManifest.yaml
@@ -7,7 +7,7 @@ metadata:
   description: "Generic llama.cpp engine base. Pick any GGUF model at install via env."
   appid: llamacppllmbasev3
   title: llama.cpp Engine Base
-  version: '1.2.37'
+  version: '1.2.47'
   categories:
     - AI
     - models
@@ -36,37 +36,10 @@ workloadReplicas:
   llminit: 1
 spec:
   onlyAdmin: true
-  versionName: 'b10588'
+  versionName: 'b10969 / b10731-rocm / b10752-intel / b10884-intel-iGPU'
   upgradeDescription: |
-    Chart 1.2.37 — bump engine to `server-cuda12-b10588`.
+    Chart 1.2.47 — Upgrade the NVIDIA / Spark CUDA image from server-cuda12-b10752 to server-cuda12-b10969.
 
-    **Images**
-    - llama.cpp: `docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10454` → `server-cuda12-b10588` (amd64 + arm64).
-    - llm-init: unchanged at `v1.5.0`.
-
-    134 upstream commits; core libraries move to llama.cpp 0.2.0 / ggml 0.21.0.
-
-    **New model architectures** — clone this base and point `MODEL_SOURCE` / `MODEL_NAME` at the GGUF:
-    - BailingMoE3 ([#26608](https://github.com/ggml-org/llama.cpp/pull/26608)), with DSpark speculative decode ([#27508](https://github.com/ggml-org/llama.cpp/pull/27508))
-    - Granite SWA / GraniteMoE SWA ([#25505](https://github.com/ggml-org/llama.cpp/pull/25505))
-    - dots3-note, including vision + audio through `mmproj` ([#27060](https://github.com/ggml-org/llama.cpp/pull/27060), [#27524](https://github.com/ggml-org/llama.cpp/pull/27524))
-    - DSpark drafters for LFM2 ([#27383](https://github.com/ggml-org/llama.cpp/pull/27383)) and speculators-format checkpoints ([#26275](https://github.com/ggml-org/llama.cpp/pull/26275))
-    - Nemotron 3 Ultra conversion fix ([#27101](https://github.com/ggml-org/llama.cpp/pull/27101))
-
-    **GPU and performance**
-    - DGX Spark (`nvidia-gb10`): faster batch-size-1 decode for dense models via MMVQ `nwarps=8` ([#26843](https://github.com/ggml-org/llama.cpp/pull/26843))
-    - CUDA MMVQ→MMQ decode crossover is now tuned per GPU and quant type ([#26079](https://github.com/ggml-org/llama.cpp/pull/26079)); cuBLAS handles use a static workspace ([#26574](https://github.com/ggml-org/llama.cpp/pull/26574))
-    - Layer auto-fit accounts for `n_streams` ([#27496](https://github.com/ggml-org/llama.cpp/pull/27496)); V is built as a view of K in iSWA attention ([#27392](https://github.com/ggml-org/llama.cpp/pull/27392))
-    - Fixes: backend split scheduler race ([#26040](https://github.com/ggml-org/llama.cpp/pull/26040)) and a null-pointer deref in speculative decode ([#27404](https://github.com/ggml-org/llama.cpp/pull/27404))
-
-    **Server and API**
-    - `/metrics` and `/slots` stay reachable while the server is sleeping ([#27376](https://github.com/ggml-org/llama.cpp/pull/27376))
-    - Model endpoints become private when `--api-key` authentication is enabled ([#26347](https://github.com/ggml-org/llama.cpp/pull/26347)); this chart runs without an engine API key, so the internal entrance is unaffected
-    - `--mmproj-device` places the vision projector on a chosen device ([#23255](https://github.com/ggml-org/llama.cpp/pull/23255)); WebP inputs are accepted via ffmpeg ([#27520](https://github.com/ggml-org/llama.cpp/pull/27520))
-    - Embedding output combined with draft-MTP is fixed ([#27400](https://github.com/ggml-org/llama.cpp/pull/27400)) — relevant to `MODEL_MODE=embedding`
-    - mtmd fixes for LFM2 image tiling, DeepSeek-OCR and Granite preprocessing; refreshed built-in WebUI
-
-    See https://github.com/ggml-org/llama.cpp/compare/b10454...b10588 for the full upstream diff.
   fullDescription: |
     **IMPORTANT NOTE**  
     This app is a template and cannot be used on its own. To use it, set the environment variables below to connect a specific model.
@@ -78,7 +51,7 @@ spec:
     - `MODEL_SUPPORTS`: Model capability options (Vision / Tools / Thinking / None). Pick what your model supports. This is required.
     - `ENGINE_ARGS`: `-c 8192 -ngl all -fa on`. Extra llama.cpp options, like `-c 8192 -ngl all -fa on`. To fully use GPU, add `-ngl all`. You can also use values like `LLAMA_ARG_*=val`.
     - `LOG_LEVEL`: `debug`, `info`, `warn`, or `error`. Log level: choose from `debug`, `info`, `warn`, or `error`.
-    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192`, or `8192Mi`. Model memory needed. Accepts values like `8Gi`, `8192`, or `8192Mi`. On NVIDIA, this sets `nvidia.com/gpumem` in MiB; on Spark, sets pod memory.
+    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192`, or `8192Mi`. Model memory needed. Accepts values like `8Gi`, `8192`, or `8192Mi`. On NVIDIA, this sets `nvidia.com/gpumem` in MiB; on Spark and `intel` integrated GPUs, sets pod memory (include model, KV cache and runtime overhead); on `amd-gpu` and `intel-gpu` declares the discrete VRAM quota through the `gpu.bytetrade.io/required-gpu-memory` pod annotation.
     
     **How It Works**
     First, `llm-init` downloads your model and gets it ready, then `llama.cpp` starts serving the model when it’s ready. These two parts run separately but share files and progress info.
@@ -92,6 +65,9 @@ spec:
     |----------------|-------------------|---------------------------------------------------------------------|
     | `nvidia`       | GPU (CUDA)        | Set `LLAMACPP_REQUIRED_GPU_MEMORY`; use `-ngl all` for full GPU use |
     | `nvidia-gb10`  | GPU (Spark)       | Like above, but memory is managed by the pod                        |
+    | `amd-gpu`      | GPU (ROCm)        | Discrete AMD (e.g. Radeon AI PRO R9700 / gfx1201). Image `server-rocm-b10731`. Use `-ngl all`. Do not use `amd` (Ryzen iGPU). |
+    | `intel` | Intel integrated GPU (Panther Lake / SYCL) | `server-intel-b10884`; `-ngl all`; LLAMACPP_REQUIRED_GPU_MEMORY = shared pod memory (model + KV + runtime). |
+    | `intel-gpu`    | GPU (SYCL)        | Discrete Intel Arc (e.g. Arc Pro B70). Image `server-intel-b10752`. Use `-ngl all`. |
 
     **Model Storage**
     All model files are stored in a shared folder at `appCommon/huggingface`, mounted as `/cache/hf/hub`. All similar apps use this location for models.
@@ -107,13 +83,23 @@ spec:
   locale:
     - en-US
     - zh-CN
+    - de-DE
+    - es-ES
+    - it-IT
+    - fr-FR
+    - ja-JP
   license:
     - text: MIT
       url: https://github.com/ggml-org/llama.cpp?tab=MIT-1-ov-file
 
   accelerator:
-    #   nvidia       — discrete GPU; LLAMACPP_REQUIRED_GPU_MEMORY -> gpumem
-    #   nvidia-gb10  — DGX Spark unified memory; same env -> pod memory
+    #   nvidia       - discrete GPU; LLAMACPP_REQUIRED_GPU_MEMORY -> nvidia.com/gpumem
+    #   nvidia-gb10  - DGX Spark unified memory; same env -> pod memory
+    #   amd-gpu      - discrete AMD ROCm; LLAMACPP_REQUIRED_GPU_MEMORY -> gpu-memory pod annotation
+    #   intel-gpu    - discrete Intel Arc SYCL; LLAMACPP_REQUIRED_GPU_MEMORY -> gpu-memory pod annotation
+    # The two discrete non-NVIDIA modes have no GPU-memory extended resource to
+    # carry the quota, so the chart puts it on gpu.bytetrade.io/required-gpu-memory
+    # and app-service reads that to resolve requiredGPUMemory: "-1".
     - mode: nvidia
       limitedCpu: "-1"
       requiredCpu: "-1"
@@ -130,11 +116,37 @@ spec:
       limitedDisk: 500Gi
       limitedMemory: "-1"
       requiredMemory: "-1"
+    - mode: amd-gpu
+      limitedCpu: "-1"
+      requiredCpu: "-1"
+      requiredDisk: 50Mi
+      limitedDisk: 500Gi
+      limitedMemory: "-1"
+      requiredMemory: "-1"
+      requiredGPUMemory: "-1"
+      limitedGPUMemory: "-1"
+    - mode: intel
+      limitedCpu: "-1"
+      requiredCpu: "-1"
+      requiredDisk: 50Mi
+      limitedDisk: 500Gi
+      limitedMemory: "-1"
+      requiredMemory: "-1"
+    - mode: intel-gpu
+      limitedCpu: "-1"
+      requiredCpu: "-1"
+      requiredDisk: 50Mi
+      limitedDisk: 500Gi
+      limitedMemory: "-1"
+      requiredMemory: "-1"
+      requiredGPUMemory: "-1"
+      limitedGPUMemory: "-1"
 
   supportArch:
     - amd64
     - arm64
 permission:
+  appCache: true
   appData: true
   appCommon: true
 options:
@@ -197,7 +209,7 @@ envs:
     type: string
     editable: true
     applyOnChange: true
-    description:  "Example: -c 8192 -ngl all -fa on. Extra llama.cpp options, to fully use GPU, add `-ngl all`. You can also use values like `LLAMA_ARG_*=val`."
+    description:  "Example: -c 8192 -ngl all -fa on. Extra llama.cpp options, to fully use GPU, add `-ngl all`. You can also use values like `LLAMA_ARG_*=val`. Note that `-c` sizes the whole KV cache rather than one request, and omitting `-np` selects llama.cpp's default of 4 slots sharing it — set `-np 1` for a single long conversation, or raise `-c` to match the concurrency you want."
   - envName: LOG_LEVEL           # single-select: debug | info | warn | error
     required: false
     type: string
@@ -252,5 +264,5 @@ envs:
     type: string
     editable: false
     applyOnChange: false
-    description:  "Example: `8Gi`, `8192`, or `8192Mi`. GPU memory required for the model. On NVIDIA this sets `nvidia.com/gpumem` in MiB. On Spark it sets pod memory."
+    description:  "Example: `8Gi`, `8192`, or `8192Mi`. GPU memory required for the model. On NVIDIA this sets `nvidia.com/gpumem` in MiB. On Spark and intel integrated GPUs it sets pod memory; include model, KV cache and runtime overhead. On amd-gpu and intel-gpu it declares the discrete VRAM quota through the gpu.bytetrade.io/required-gpu-memory pod annotation."
     regex: "^[0-9]+(Gi|Mi)?$"
diff --git a/llamacppllmbasev3/i18n/de-DE/OlaresManifest.yaml b/llamacppllmbasev3/i18n/de-DE/OlaresManifest.yaml
new file mode 100644
index 000000000..796ca5739
--- /dev/null
+++ b/llamacppllmbasev3/i18n/de-DE/OlaresManifest.yaml
@@ -0,0 +1,40 @@
+metadata:
+  description: "Generische llama.cpp-Engine-Basis. Beliebiges GGUF-Modell bei der Installation per Env wählen"
+  title: llama.cpp Engine Base
+
+spec:
+  upgradeDescription: |
+    Chart 1.2.47 — NVIDIA-/Spark-CUDA-Image von server-cuda12-b10752 auf server-cuda12-b10969 aktualisiert.
+
+  fullDescription: |
+    **WICHTIGER HINWEIS**  
+    Diese App ist eine Vorlage und kann nicht allein genutzt werden. Setzen Sie die unten stehenden Umgebungsvariablen, um ein konkretes Modell anzubinden.
+
+    **Einrichtungsleitfaden für Umgebungsvariablen**
+    - `MODEL_SOURCE`: `hf://<repo> --include <file>.gguf`. Quelle zum Herunterladen Ihres Modells. Bei Split-Dateien nur den ersten Teil angeben; bei Vision-Language-Modellen den Projektor per Komma hinzufügen, z. B. `hf://<repo> --include model.gguf,hf://<repo> --include mmproj-F16.gguf`.
+    - `MODEL_NAME`: `owner/repo:quant`. Name Ihres Modells im Format `owner/repo:quant`, z. B. `unsloth/Qwen3.5-2B-GGUF:UD-Q4_K_XL`. Wird auch als OpenAI-Alias verwendet.
+    - `MODEL_MODE`: `chat` oder `embedding`. Bei embedding fügt das Chart automatisch `--embedding` hinzu.
+    - `MODEL_SUPPORTS`: Modellfähigkeiten (Vision / Tools / Thinking / None). Wählen Sie, was Ihr Modell unterstützt. Erforderlich.
+    - `ENGINE_ARGS`: `-c 8192 -ngl all -fa on`. Zusätzliche llama.cpp-Optionen wie `-c 8192 -ngl all -fa on`. Für volle GPU-Nutzung `-ngl all` hinzufügen. Auch Werte wie `LLAMA_ARG_*=val` möglich.
+    - `LOG_LEVEL`: `debug`, `info`, `warn` oder `error`. Log-Level: Wählen Sie zwischen `debug`, `info`, `warn` oder `error`.
+    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192` oder `8192Mi`. Benötigter Modellspeicher. Akzeptiert Werte wie `8Gi`, `8192` oder `8192Mi`. Auf NVIDIA setzt dies `nvidia.com/gpumem` in MiB; auf Spark die Pod-Memory; auf `amd-gpu` und `intel-gpu` deklariert es das diskrete VRAM-Kontingent über die Pod-Annotation `gpu.bytetrade.io/required-gpu-memory`.
+
+    **Funktionsweise**
+    Zuerst lädt `llm-init` Ihr Modell herunter und bereitet es vor, dann startet `llama.cpp` den Serving, wenn es bereit ist. Beide Teile laufen getrennt, teilen aber Dateien und Fortschrittsinformationen.
+
+    **Stack**
+    - **llama.cpp**: Führt das Modell aus und bedient API-Anfragen (Port 8081)
+    - **llm-init**: Lädt das Modell herunter, zeigt Fortschritt und stellt die OpenAI-API auf Port 8090 bereit, mit Proxy zu `llamacpp`.
+
+    **Beschleuniger-Modi**
+    | Modus          | Gerät             | Details                                                            |
+    |----------------|-------------------|---------------------------------------------------------------------|
+    | `nvidia`       | GPU (CUDA)        | `LLAMACPP_REQUIRED_GPU_MEMORY` setzen; `-ngl all` für volle GPU-Nutzung |
+    | `nvidia-gb10`  | GPU (Spark)       | Wie oben, aber Speicher wird vom Pod verwaltet                        |
+
+    **Modellspeicher**
+    Alle Modelldateien liegen im gemeinsamen Ordner `appCommon/huggingface`, gemountet als `/cache/hf/hub`. Ähnliche Apps nutzen denselben Ort für Modelle.
+
+    **Ressourcentipps**
+    - CPU, RAM und GPU konfigurieren sich automatisch anhand des Charts; Disk-Einstellungen sind fest.
+    - Für die GPU `-ngl all` in `ENGINE_ARGS` setzen. `-ngl all` für volles GPU-Offload.
diff --git a/llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml b/llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml
index 835e2e842..e958e8d74 100644
--- a/llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml
+++ b/llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml
@@ -4,25 +4,8 @@ metadata:
 
 spec:
   upgradeDescription: |
-    Chart 1.2.33 — bump engine to `server-cuda12-b10454` and llm-init to `v1.3.8`.
+    Chart 1.2.47 — Upgrade the NVIDIA / Spark CUDA image from server-cuda12-b10752 to server-cuda12-b10969.
 
-    **Images**
-    - llama.cpp: `docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10362` → `server-cuda12-b10454` (latest published beclab CUDA 12 server tag; amd64 + arm64). Upstream GitHub is already at b10456; no beclab b10455/b10456 image yet.
-    - llm-init: `docker.io/beclab/llm-init:v1.3.5` → `v1.3.8`.
-
-    **Muse Glimmer 30B (Meta)** — requires **b10353+**; **b10454 includes** the `muse-glimmer` architecture ([llama.cpp #26841](https://github.com/ggml-org/llama.cpp/pull/26841)). Older builds (e.g. b10331) fail with `unknown model architecture: 'muse-glimmer'`.
-
-    Clone this base and point env at [meta-models/Muse-Glimmer-30B-GGUF](https://huggingface.co/meta-models/Muse-Glimmer-30B-GGUF) or community quants (e.g. unsloth). Typical layout:
-    - Main GGUF + `mmproj` in `MODEL_SOURCE` (comma-separated) for vision
-    - DFlash drafter (`dflash-kquant.gguf`) + `--spec-type draft-dflash` or `draft-dspark` in `ENGINE_ARGS` for speculative decode
-
-    **Also in b10362 → b10454:**
-    - Chat templates receive OpenAI `reasoning_effort` ([llama.cpp #26941](https://github.com/ggml-org/llama.cpp/pull/26941)); Qwen3.8 uses `xhigh` / `medium` / `low` via `--chat-template-kwargs` or the request body
-    - MTP draft type auto-detect; DFlash / DSpark backend sampling
-    - MiniMax Text01 / M1 conversion support
-    - Server `/metrics` and `/slots` stay available during decode; mtmd and jinja fixes
-
-    See https://github.com/ggml-org/llama.cpp/releases for full upstream release notes.
   fullDescription: |
     **IMPORTANT NOTE**  
     This app is a template and cannot be used on its own. To use it, set the environment variables below to connect a specific model.
@@ -34,7 +17,7 @@ spec:
     - `MODEL_SUPPORTS`: Model capability options (Vision / Tools / Thinking / None). Pick what your model supports. This is required.
     - `ENGINE_ARGS`: `-c 8192 -ngl all -fa on`. Extra llama.cpp options, like `-c 8192 -ngl all -fa on`. To fully use GPU, add `-ngl all`. You can also use values like `LLAMA_ARG_*=val`.
     - `LOG_LEVEL`: `debug`, `info`, `warn`, or `error`. Log level: choose from `debug`, `info`, `warn`, or `error`.
-    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192`, or `8192Mi`. Model memory needed. Accepts values like `8Gi`, `8192`, or `8192Mi`. On NVIDIA, this sets `nvidia.com/gpumem` in MiB; on Spark, sets pod memory.
+    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192`, or `8192Mi`. Model memory needed. Accepts values like `8Gi`, `8192`, or `8192Mi`. On NVIDIA, this sets `nvidia.com/gpumem` in MiB; on Spark, sets pod memory; on `amd-gpu` and `intel-gpu` declares the discrete VRAM quota through the `gpu.bytetrade.io/required-gpu-memory` pod annotation.
     
     **How It Works**
     First, `llm-init` downloads your model and gets it ready, then `llama.cpp` starts serving the model when it’s ready. These two parts run separately but share files and progress info.
@@ -48,6 +31,9 @@ spec:
     |----------------|-------------------|---------------------------------------------------------------------|
     | `nvidia`       | GPU (CUDA)        | Set `LLAMACPP_REQUIRED_GPU_MEMORY`; use `-ngl all` for full GPU use |
     | `nvidia-gb10`  | GPU (Spark)       | Like above, but memory is managed by the pod                        |
+    | `amd-gpu`      | GPU (ROCm)        | Discrete AMD (R9700 / gfx1201). Image `server-rocm-b10731`. Do not use `amd`. |
+    | `intel` | Intel integrated GPU (Panther Lake / SYCL) | `server-intel-b10884`; `-ngl all`; LLAMACPP_REQUIRED_GPU_MEMORY = shared pod memory (model + KV + runtime). |
+    | `intel-gpu`    | GPU (SYCL)        | Discrete Intel Arc (e.g. Arc Pro B70). Image `server-intel-b10752`. Use `-ngl all`. |
 
     **Model Storage**
     All model files are stored in a shared folder at `appCommon/huggingface`, mounted as `/cache/hf/hub`. All similar apps use this location for models.
diff --git a/llamacppllmbasev3/i18n/es-ES/OlaresManifest.yaml b/llamacppllmbasev3/i18n/es-ES/OlaresManifest.yaml
new file mode 100644
index 000000000..a3f1b8340
--- /dev/null
+++ b/llamacppllmbasev3/i18n/es-ES/OlaresManifest.yaml
@@ -0,0 +1,40 @@
+metadata:
+  description: "Base genérica del motor llama.cpp. Elija cualquier modelo GGUF en la instalación vía env"
+  title: llama.cpp Engine Base
+
+spec:
+  upgradeDescription: |
+    Chart 1.2.47 — Actualiza la imagen CUDA de NVIDIA / Spark de server-cuda12-b10752 a server-cuda12-b10969.
+
+  fullDescription: |
+    **NOTA IMPORTANTE**  
+    Esta app es una plantilla y no puede usarse sola. Para usarla, configure las variables de entorno siguientes para conectar un modelo concreto.
+
+    **Guía de configuración de variables de entorno**
+    - `MODEL_SOURCE`: `hf://<repo> --include <file>.gguf`. Fuente para descargar su modelo. En archivos partidos, incluya solo la primera parte; en modelos visión-lenguaje, añada el proyector con una coma, p. ej. `hf://<repo> --include model.gguf,hf://<repo> --include mmproj-F16.gguf`.
+    - `MODEL_NAME`: `owner/repo:quant`. Nombre de su modelo en formato `owner/repo:quant`, p. ej. `unsloth/Qwen3.5-2B-GGUF:UD-Q4_K_XL`. También se usa como alias OpenAI.
+    - `MODEL_MODE`: `chat` o `embedding`. Si elige embedding, el chart añade `--embedding` automáticamente.
+    - `MODEL_SUPPORTS`: Capacidades del modelo (Vision / Tools / Thinking / None). Elija lo que su modelo admite. Obligatorio.
+    - `ENGINE_ARGS`: `-c 8192 -ngl all -fa on`. Opciones extra de llama.cpp, como `-c 8192 -ngl all -fa on`. Para usar la GPU al completo, añada `-ngl all`. También valores como `LLAMA_ARG_*=val`.
+    - `LOG_LEVEL`: `debug`, `info`, `warn` o `error`. Nivel de log: elija entre `debug`, `info`, `warn` o `error`.
+    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192` o `8192Mi`. Memoria de modelo necesaria. Admite valores como `8Gi`, `8192` o `8192Mi`. En NVIDIA establece `nvidia.com/gpumem` en MiB; en Spark, la memoria del pod; en `amd-gpu` e `intel-gpu` declara la cuota de VRAM discreta mediante la anotación de pod `gpu.bytetrade.io/required-gpu-memory`.
+
+    **Cómo funciona**
+    Primero, `llm-init` descarga su modelo y lo prepara; luego `llama.cpp` empieza a servir el modelo cuando está listo. Ambas partes se ejecutan por separado pero comparten archivos e información de progreso.
+
+    **Stack**
+    - **llama.cpp**: Ejecuta el modelo y atiende peticiones API (puerto 8081)
+    - **llm-init**: Descarga el modelo, muestra progreso y proporciona la API OpenAI en el puerto 8090, haciendo proxy a `llamacpp`.
+
+    **Modos de acelerador**
+    | Modo           | Dispositivo       | Detalles                                                           |
+    |----------------|-------------------|---------------------------------------------------------------------|
+    | `nvidia`       | GPU (CUDA)        | Establezca `LLAMACPP_REQUIRED_GPU_MEMORY`; use `-ngl all` para GPU completa |
+    | `nvidia-gb10`  | GPU (Spark)       | Como arriba, pero la memoria la gestiona el pod                       |
+
+    **Almacenamiento de modelos**
+    Todos los archivos de modelo se guardan en la carpeta compartida `appCommon/huggingface`, montada como `/cache/hf/hub`. Apps similares usan esta ubicación para modelos.
+
+    **Consejos de recursos**
+    - CPU, RAM y GPU se autoconfiguran según el chart; los ajustes de disco son fijos.
+    - Para usar la GPU, establezca `-ngl all` en `ENGINE_ARGS`. Use `-ngl all` para offload GPU completo.
diff --git a/llamacppllmbasev3/i18n/fr-FR/OlaresManifest.yaml b/llamacppllmbasev3/i18n/fr-FR/OlaresManifest.yaml
new file mode 100644
index 000000000..00d610e5b
--- /dev/null
+++ b/llamacppllmbasev3/i18n/fr-FR/OlaresManifest.yaml
@@ -0,0 +1,40 @@
+metadata:
+  description: "Base générique du moteur llama.cpp. Choisissez n'importe quel modèle GGUF à l'installation via env"
+  title: llama.cpp Engine Base
+
+spec:
+  upgradeDescription: |
+    Chart 1.2.47 — Mise à jour CUDA NVIDIA / Spark de server-cuda12-b10752 vers server-cuda12-b10969.
+
+  fullDescription: |
+    **NOTE IMPORTANTE**  
+    Cette app est un modèle et ne peut pas être utilisée seule. Pour l'utiliser, définissez les variables d'environnement ci-dessous pour connecter un modèle spécifique.
+
+    **Guide de configuration des variables d'environnement**
+    - `MODEL_SOURCE` : `hf://<repo> --include <file>.gguf`. Source pour télécharger votre modèle. Pour les fichiers fractionnés, n'incluez que la première partie ; pour les modèles vision-langage, ajoutez le projecteur avec une virgule, ex. `hf://<repo> --include model.gguf,hf://<repo> --include mmproj-F16.gguf`.
+    - `MODEL_NAME` : `owner/repo:quant`. Nom de votre modèle au format `owner/repo:quant`, ex. `unsloth/Qwen3.5-2B-GGUF:UD-Q4_K_XL`. Aussi utilisé comme alias OpenAI.
+    - `MODEL_MODE` : `chat` ou `embedding`. Si vous choisissez embedding, le chart ajoute automatiquement `--embedding`.
+    - `MODEL_SUPPORTS` : Capacités du modèle (Vision / Tools / Thinking / None). Choisissez ce que votre modèle prend en charge. Obligatoire.
+    - `ENGINE_ARGS` : `-c 8192 -ngl all -fa on`. Options llama.cpp supplémentaires, comme `-c 8192 -ngl all -fa on`. Pour utiliser pleinement le GPU, ajoutez `-ngl all`. Aussi des valeurs comme `LLAMA_ARG_*=val`.
+    - `LOG_LEVEL` : `debug`, `info`, `warn` ou `error`. Niveau de log : choisissez parmi `debug`, `info`, `warn` ou `error`.
+    - `LLAMACPP_REQUIRED_GPU_MEMORY` : `8Gi`, `8192` ou `8192Mi`. Mémoire modèle nécessaire. Accepte des valeurs comme `8Gi`, `8192` ou `8192Mi`. Sur NVIDIA, définit `nvidia.com/gpumem` en MiB ; sur Spark, la mémoire du pod ; sur `amd-gpu` et `intel-gpu`, déclare le quota de VRAM discrète via l'annotation de pod `gpu.bytetrade.io/required-gpu-memory`.
+
+    **Fonctionnement**
+    D'abord, `llm-init` télécharge votre modèle et le prépare, puis `llama.cpp` commence à servir le modèle quand il est prêt. Les deux parties tournent séparément mais partagent fichiers et infos de progression.
+
+    **Stack**
+    - **llama.cpp** : Exécute le modèle et sert les requêtes API (port 8081)
+    - **llm-init** : Télécharge le modèle, affiche la progression et fournit l'API OpenAI sur le port 8090, en proxy vers `llamacpp`.
+
+    **Modes accélérateur**
+    | Mode           | Périphérique      | Détails                                                            |
+    |----------------|-------------------|---------------------------------------------------------------------|
+    | `nvidia`       | GPU (CUDA)        | Définir `LLAMACPP_REQUIRED_GPU_MEMORY` ; utiliser `-ngl all` pour GPU complet |
+    | `nvidia-gb10`  | GPU (Spark)       | Comme ci-dessus, mais la mémoire est gérée par le pod                 |
+
+    **Stockage des modèles**
+    Tous les fichiers modèle sont stockés dans le dossier partagé `appCommon/huggingface`, monté comme `/cache/hf/hub`. Des apps similaires utilisent cet emplacement pour les modèles.
+
+    **Conseils ressources**
+    - CPU, RAM et GPU s'autoconfigurent selon le chart ; les réglages disque sont fixes.
+    - Pour utiliser le GPU, définissez `-ngl all` dans `ENGINE_ARGS`. Utilisez `-ngl all` pour un offload GPU complet.
diff --git a/llamacppllmbasev3/i18n/it-IT/OlaresManifest.yaml b/llamacppllmbasev3/i18n/it-IT/OlaresManifest.yaml
new file mode 100644
index 000000000..c65b2eb9c
--- /dev/null
+++ b/llamacppllmbasev3/i18n/it-IT/OlaresManifest.yaml
@@ -0,0 +1,40 @@
+metadata:
+  description: "Base generica del motore llama.cpp. Scegli qualsiasi modello GGUF in installazione via env"
+  title: llama.cpp Engine Base
+
+spec:
+  upgradeDescription: |
+    Chart 1.2.47 — Aggiornamento CUDA NVIDIA / Spark da server-cuda12-b10752 a server-cuda12-b10969.
+
+  fullDescription: |
+    **NOTA IMPORTANTE**  
+    Questa app è un template e non può essere usata da sola. Per usarla, imposta le variabili d'ambiente sotto per collegare un modello specifico.
+
+    **Guida alla configurazione delle variabili d'ambiente**
+    - `MODEL_SOURCE`: `hf://<repo> --include <file>.gguf`. Sorgente da cui scaricare il modello. Per file suddivisi includi solo la prima parte; per modelli vision-language aggiungi il proiettore con una virgola, es. `hf://<repo> --include model.gguf,hf://<repo> --include mmproj-F16.gguf`.
+    - `MODEL_NAME`: `owner/repo:quant`. Nome del modello in formato `owner/repo:quant`, es. `unsloth/Qwen3.5-2B-GGUF:UD-Q4_K_XL`. Usato anche come alias OpenAI.
+    - `MODEL_MODE`: `chat` o `embedding`. Se scegli embedding, il chart aggiunge automaticamente `--embedding`.
+    - `MODEL_SUPPORTS`: Capacità del modello (Vision / Tools / Thinking / None). Scegli ciò che il modello supporta. Obbligatorio.
+    - `ENGINE_ARGS`: `-c 8192 -ngl all -fa on`. Opzioni extra llama.cpp, come `-c 8192 -ngl all -fa on`. Per usare appieno la GPU, aggiungi `-ngl all`. Anche valori come `LLAMA_ARG_*=val`.
+    - `LOG_LEVEL`: `debug`, `info`, `warn` o `error`. Livello di log: scegli tra `debug`, `info`, `warn` o `error`.
+    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192` o `8192Mi`. Memoria modello necessaria. Accetta valori come `8Gi`, `8192` o `8192Mi`. Su NVIDIA imposta `nvidia.com/gpumem` in MiB; su Spark, la memoria del pod; su `amd-gpu` e `intel-gpu` dichiara la quota di VRAM discreta tramite l'annotazione del pod `gpu.bytetrade.io/required-gpu-memory`.
+
+    **Come funziona**
+    Prima `llm-init` scarica il modello e lo prepara, poi `llama.cpp` inizia a servirlo quando è pronto. Le due parti girano separatemente ma condividono file e info di progresso.
+
+    **Stack**
+    - **llama.cpp**: Esegue il modello e serve le richieste API (porta 8081)
+    - **llm-init**: Scarica il modello, mostra il progresso e fornisce l'API OpenAI sulla porta 8090, facendo proxy verso `llamacpp`.
+
+    **Modalità acceleratore**
+    | Modalità       | Dispositivo       | Dettagli                                                           |
+    |----------------|-------------------|---------------------------------------------------------------------|
+    | `nvidia`       | GPU (CUDA)        | Imposta `LLAMACPP_REQUIRED_GPU_MEMORY`; usa `-ngl all` per GPU piena |
+    | `nvidia-gb10`  | GPU (Spark)       | Come sopra, ma la memoria è gestita dal pod                          |
+
+    **Archiviazione modelli**
+    Tutti i file modello sono nella cartella condivisa `appCommon/huggingface`, montata come `/cache/hf/hub`. App simili usano questa posizione per i modelli.
+
+    **Suggerimenti sulle risorse**
+    - CPU, RAM e GPU si autoconfigurano in base al chart; le impostazioni disco sono fisse.
+    - Per usare la GPU, imposta `-ngl all` in `ENGINE_ARGS`. Usa `-ngl all` per offload GPU completo.
diff --git a/llamacppllmbasev3/i18n/ja-JP/OlaresManifest.yaml b/llamacppllmbasev3/i18n/ja-JP/OlaresManifest.yaml
new file mode 100644
index 000000000..c70f4d8a4
--- /dev/null
+++ b/llamacppllmbasev3/i18n/ja-JP/OlaresManifest.yaml
@@ -0,0 +1,40 @@
+metadata:
+  description: "汎用 llama.cpp エンジンベース。インストール時に env で任意の GGUF モデルを選択"
+  title: llama.cpp Engine Base
+
+spec:
+  upgradeDescription: |
+    Chart 1.2.47 — NVIDIA / Spark CUDA イメージを server-cuda12-b10752 から server-cuda12-b10969 に更新。
+
+  fullDescription: |
+    **重要なお知らせ**  
+    このアプリはテンプレートであり、単体では使えません。下の環境変数を設定して特定モデルを接続してください。
+
+    **環境変数セットアップガイド**
+    - `MODEL_SOURCE`: `hf://<repo> --include <file>.gguf`。モデルのダウンロード元。分割ファイルは先頭パートのみ; ビジョン言語モデルはカンマでプロジェクタを追加。例: `hf://<repo> --include model.gguf,hf://<repo> --include mmproj-F16.gguf`。
+    - `MODEL_NAME`: `owner/repo:quant`。`owner/repo:quant` 形式のモデル名。例: `unsloth/Qwen3.5-2B-GGUF:UD-Q4_K_XL`。OpenAI エイリアスとしても使用。
+    - `MODEL_MODE`: `chat` または `embedding`。embedding を選ぶとチャートが自動で `--embedding` を追加。
+    - `MODEL_SUPPORTS`: モデル能力（Vision / Tools / Thinking / None）。モデルが対応するものを選択。必須。
+    - `ENGINE_ARGS`: `-c 8192 -ngl all -fa on`。`-c 8192 -ngl all -fa on` などの追加 llama.cpp オプション。GPU をフルに使うには `-ngl all`。`LLAMA_ARG_*=val` 形式も可。
+    - `LOG_LEVEL`: `debug`、`info`、`warn`、または `error`。ログレベルは `debug`、`info`、`warn`、`error` から選択。
+    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`、`8192`、または `8192Mi`。必要なモデルメモリ。`8Gi`、`8192`、`8192Mi` などの値を使用可能。NVIDIA では MiB 単位の `nvidia.com/gpumem` を設定; Spark ではポッドメモリを設定; `amd-gpu` と `intel-gpu` では Pod アノテーション `gpu.bytetrade.io/required-gpu-memory` でディスクリート VRAM クォータを宣言。
+
+    **動作の仕組み**
+    まず `llm-init` がモデルをダウンロードして準備し、準備ができたら `llama.cpp` がサービングを開始します。両者は別々に動作しますが、ファイルと進捗情報を共有します。
+
+    **スタック**
+    - **llama.cpp**: モデルを実行し API リクエストを処理（ポート 8081）
+    - **llm-init**: モデルをダウンロードし、進捗を表示し、ポート 8090 で OpenAI API を提供して `llamacpp` へプロキシ。
+
+    **アクセラレータモード**
+    | モード         | デバイス          | 詳細                                                               |
+    |----------------|-------------------|---------------------------------------------------------------------|
+    | `nvidia`       | GPU (CUDA)        | `LLAMACPP_REQUIRED_GPU_MEMORY` を設定; フル GPU には `-ngl all`     |
+    | `nvidia-gb10`  | GPU (Spark)       | 上記と同様だがメモリはポッドが管理                                    |
+
+    **モデルストレージ**
+    すべてのモデルファイルは共有フォルダ `appCommon/huggingface` に保存され、`/cache/hf/hub` にマウントされます。類似アプリもモデルにこの場所を使います。
+
+    **リソースのヒント**
+    - CPU、RAM、GPU はチャートに基づき自動設定; ディスク設定は固定。
+    - GPU を使うには `ENGINE_ARGS` に `-ngl all` を設定。フル GPU オフロードには `-ngl all`。
diff --git a/llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml b/llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml
index 666700a95..85b5dba5e 100644
--- a/llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml
+++ b/llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml
@@ -4,25 +4,8 @@ metadata:
 
 spec:
   upgradeDescription: |
-    Chart 1.2.33 — 引擎升级到 `server-cuda12-b10454`，llm-init 升级到 `v1.3.8`。
+    Chart 1.2.47 — NVIDIA / Spark CUDA 镜像从 server-cuda12-b10752 升级到 server-cuda12-b10969。
 
-    **镜像**
-    - llama.cpp：`docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10362` → `server-cuda12-b10454`（当前已发布的最新 beclab CUDA 12 server 标签；amd64 + arm64）。上游 GitHub 已到 b10456；尚无 beclab 的 b10455/b10456 镜像。
-    - llm-init：`docker.io/beclab/llm-init:v1.3.5` → `v1.3.8`。
-
-    **Muse Glimmer 30B（Meta）** — 需要 **b10353+**；**b10454 已包含** `muse-glimmer` 架构（[llama.cpp #26841](https://github.com/ggml-org/llama.cpp/pull/26841)）。更旧的构建（如 b10331）会报 `unknown model architecture: 'muse-glimmer'`。
-
-    克隆此基座后，将环境变量指向 [meta-models/Muse-Glimmer-30B-GGUF](https://huggingface.co/meta-models/Muse-Glimmer-30B-GGUF) 或社区量化（如 unsloth）。常见布局：
-    - 主 GGUF + `mmproj` 写入 `MODEL_SOURCE`（逗号分隔）以支持视觉
-    - DFlash 草稿模型（`dflash-kquant.gguf`）并在 `ENGINE_ARGS` 中加 `--spec-type draft-dflash` 或 `draft-dspark` 以开启投机解码
-
-    **b10362 → b10454 其它变化：**
-    - 聊天模板可接收 OpenAI `reasoning_effort`（[llama.cpp #26941](https://github.com/ggml-org/llama.cpp/pull/26941)）；Qwen3.8 可用 `--chat-template-kwargs` 或请求体设置 `xhigh` / `medium` / `low`
-    - MTP 草稿类型自动检测；DFlash / DSpark 后端采样
-    - 支持 MiniMax Text01 / M1 转换
-    - 解码过程中 `/metrics` 与 `/slots` 仍可用；mtmd 与 jinja 修复
-
-    完整上游发布说明见 https://github.com/ggml-org/llama.cpp/releases
   fullDescription: |
     **重要提示**  
     此应用是一个模板，不能单独使用。请按下方环境变量设置模型后，方可使用本应用提供模型服务。
@@ -34,7 +17,7 @@ spec:
     - `MODEL_SUPPORTS`：模型能力选项（Vision / Tools / Thinking / None）。请选择您的模型支持的功能，必填。
     - `ENGINE_ARGS`：`-c 8192 -ngl all -fa on`。如需完全利用 GPU，请加 `-ngl all`。也可用如 `LLAMA_ARG_*=val` 的参数。
     - `LOG_LEVEL`：`debug`、`info`、`warn` 或 `error`。日志等级，请从上述选项中选择。
-    - `LLAMACPP_REQUIRED_GPU_MEMORY`：`8Gi`、`8192` 或 `8192Mi`。模型（仅 GPU 模式）所需显存。支持 `8Gi`、`8192`、`8192Mi` 等格式；NVIDIA 下设置 `nvidia.com/gpumem`（MiB），Spark 下设置 pod 显存。
+    - `LLAMACPP_REQUIRED_GPU_MEMORY`：`8Gi`、`8192` 或 `8192Mi`。模型（仅 GPU 模式）所需显存。支持 `8Gi`、`8192`、`8192Mi` 等格式；NVIDIA 下设置 `nvidia.com/gpumem`（MiB），Spark 下设置 pod 显存，`amd-gpu` 与 `intel-gpu` 下通过 pod annotation `gpu.bytetrade.io/required-gpu-memory` 声明独立显存配额。
 
     **工作流程说明**
     首先，`llm-init` 下载并准备模型，然后 `llama.cpp` 在模型准备好后启动服务。这两个部分独立运行但共享文件与进度信息。
@@ -48,6 +31,9 @@ spec:
     |----------------|------------------|--------------------------------------------------------------|
     | `nvidia`       | GPU (CUDA)       | 需设置 `LLAMACPP_REQUIRED_GPU_MEMORY`，用 `-ngl all` 完全利用 GPU |
     | `nvidia-gb10`  | GPU (Spark)      | 类似上方，但显存由 pod 管理                                  |
+    | `amd-gpu`      | GPU (ROCm)       | AMD 独显（R9700 / gfx1201）。镜像 `server-rocm-b10731`。不要用 `amd`（核显）。 |
+    | `intel` | Intel integrated GPU (Panther Lake / SYCL) | `server-intel-b10884`; `-ngl all`; LLAMACPP_REQUIRED_GPU_MEMORY = shared pod memory (model + KV + runtime). |
+    | `intel-gpu`    | GPU (SYCL)       | Intel Arc 独显（如 Arc Pro B70）。镜像 `server-intel-b10752`。请用 `-ngl all`。 |
 
     **模型存储**
     所有模型文件存储在共享目录 `appCommon/huggingface`（挂载为 `/cache/hf/hub`），同类应用均使用此目录。
diff --git a/llamacppllmbasev3/templates/_helpers.tpl b/llamacppllmbasev3/templates/_helpers.tpl
index 3dcee99c8..71f04826f 100644
--- a/llamacppllmbasev3/templates/_helpers.tpl
+++ b/llamacppllmbasev3/templates/_helpers.tpl
@@ -23,3 +23,53 @@
 {{- $in := . -}}
 {{- trim ($in.Args | default "") -}}
 {{- end -}}
+{{- /* Olares GPU mode at install: nvidia | nvidia-gb10 | amd-gpu | intel | intel-gpu. */ -}}
+{{- define "llmbase.gpuType" -}}
+{{- $gpuObj := .Values.GPU | default dict -}}
+{{- $gpuType := .Values.gpu | default "" -}}
+{{- if not $gpuType -}}
+{{- $gpuType = $gpuObj.Type | default "nvidia" -}}
+{{- end -}}
+{{- $gpuType -}}
+{{- end -}}
+{{- /* llama.cpp engine image by accelerator. NVIDIA/Spark CUDA b10752; amd-gpu ROCm b10731; intel-gpu SYCL b10752. */ -}}
+{{- define "llamacppllmbasev3.engineImage" -}}
+{{- $gpuType := include "llmbase.gpuType" . -}}
+{{- $isGb10 := or (eq $gpuType "nvidia-gb10") (eq (include "llmbase.isGb10" .) "true") -}}
+{{- $img := .Values.engine.images | default dict -}}
+{{- if eq $gpuType "amd-gpu" -}}
+{{- $img.amdGpu | default "docker.io/beclab/ggml-org-llama.cpp:server-rocm-b10731" -}}
+{{- else if eq $gpuType "intel-gpu" -}}
+{{- $img.intelGpu | default "docker.io/beclab/ggml-org-llama.cpp:server-intel-b10752" -}}
+{{- else if eq $gpuType "intel" -}}
+{{- $img.intel | default "docker.io/beclab/ggml-org-llama.cpp:server-intel-b10884" -}}
+{{- else if $isGb10 -}}
+{{- $img.nvidiaGb10 | default "docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10969" -}}
+{{- else -}}
+{{- $img.nvidia | default "docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10969" -}}
+{{- end -}}
+{{- end -}}
+{{- /* Spark/GB10: detect from GPU.Type or node hardware (install-time .Values.nodes). */ -}}
+{{- define "llmbase.isGb10" -}}
+{{- $isGb10 := "false" -}}
+{{- if .Values.nodes -}}
+  {{- range $nodeIndex, $node := .Values.nodes -}}
+    {{- if eq $nodeIndex 0 -}}
+      {{- with $node -}}
+        {{- if .GPUS -}}
+          {{- range $gpuIndex, $gpu := .GPUS -}}
+            {{- if eq $gpuIndex 0 -}}
+              {{- with $gpu -}}
+                {{- if eq (upper .Model) "GB10" -}}
+                  {{- $isGb10 = "true" -}}
+                {{- end -}}
+              {{- end -}}
+            {{- end -}}
+          {{- end -}}
+        {{- end -}}
+      {{- end -}}
+    {{- end -}}
+  {{- end -}}
+{{- end -}}
+{{- $isGb10 -}}
+{{- end -}}
diff --git a/llamacppllmbasev3/templates/llamacpp.yaml b/llamacppllmbasev3/templates/llamacpp.yaml
index 7640a301f..141bf3f0a 100644
--- a/llamacppllmbasev3/templates/llamacpp.yaml
+++ b/llamacppllmbasev3/templates/llamacpp.yaml
@@ -13,26 +13,46 @@
 {{- $memRequest := trim ($oe.LLAMACPP_MEMORY_REQUEST | default "2Gi") -}}
 {{- $cpuLimit := trim ($oe.LLAMACPP_CPU_LIMIT | default "6") -}}
 {{- $memLimit := trim ($oe.LLAMACPP_MEMORY_LIMIT | default "35Gi") -}}
-{{- $gpuType := .Values.gpu | default "" -}}
-{{- if not $gpuType -}}
-{{- $gpuType = .Values.GPU.Type | default "nvidia" -}}
+{{- $gpuType := include "llmbase.gpuType" . -}}
+{{- $engineImage := include "llamacppllmbasev3.engineImage" . -}}
+{{- $gb10FromNodes := eq (include "llmbase.isGb10" .) "true" -}}
+{{- $isAmdGpu := eq $gpuType "amd-gpu" -}}
+{{- $isIntelGpu := eq $gpuType "intel-gpu" -}}
+{{- $isIntel := eq $gpuType "intel" -}}
+{{- $needDri := or $isAmdGpu $isIntelGpu $isIntel -}}
+{{- $unifiedMem := or $gb10FromNodes (eq $gpuType "nvidia-gb10") $isIntel -}}
+{{- /* nvidia.com/gpumem is NVIDIA-HAMi only: HAMi's extender exempts it from
+       the node capacity check, so it can carry a quota the kubelet never
+       advertises. The discrete AMD/Intel plugins advertise a card count and
+       nothing else, so their quota cannot ride on a container resource at all
+       and travels on a pod annotation instead, which is inert to the
+       scheduler. app-service reads it to resolve the requiredGPUMemory /
+       limitedGPUMemory "-1" sentinel this manifest declares for those modes;
+       without it the sentinel resolves to 0 and the card is dropped from its
+       own compute binding on resume. */ -}}
+{{- $needNvidiaGpumem := and (not $unifiedMem) (not $isAmdGpu) (not $isIntelGpu) -}}
+{{- $needGpuMemAnnotation := or $isAmdGpu $isIntelGpu -}}
+{{- $shmLimit := "16Gi" -}}
+{{- if or $isIntelGpu $isIntel -}}
+{{- $shmLimit = "10Gi" -}}
 {{- end -}}
-{{- $unifiedMem := eq $gpuType "nvidia-gb10" -}}
 {{- $engineArgs = include "llamacppllmbasev3.engineArgs" (dict "Args" $engineArgs) -}}
-{{- /* nvidia.com/gpumem needs a BARE MiB int; Spark unified memory reuses the
+{{- /* nvidia.com/gpumem needs a BARE MiB int; Spark / Intel integrated unified memory reuses the
        same value as pod memory ("<MiB>Mi"). */ -}}
 {{- $gpuMiB := include "llmbase.gpuMiB" ($oe.LLAMACPP_REQUIRED_GPU_MEMORY | default "4096") -}}
 {{- if $unifiedMem -}}
 {{- $memRequest = printf "%sMi" $gpuMiB -}}
 {{- $memLimit = printf "%sMi" $gpuMiB -}}
 {{- end -}}
-{{- $runStateHostPath := printf "%s/huggingface/llm-init-run/%s" .Values.userspace.appCommon .Release.Name -}}
+{{- $runStateHostPath := printf "%s/llm-init-run/%s" .Values.userspace.appCache .Release.Name -}}
 ---
 # llama.cpp engine Deployment. wrappers/llamacpp.sh blocks on the sentinel
 # (download ready), then execs llama-server -hf "$MODEL_NAME".
 # llm-init reverse-proxies at http://llamacpp:8081 — Service name MUST be
-# "llamacpp". GPU base: CUDA server build + gpu-inject; put `-ngl all` in
-# ENGINE_ARGS to offload all layers to the GPU.
+# "llamacpp". GPU base: CUDA (nvidia / nvidia-gb10), ROCm (amd-gpu), or
+# Intel SYCL (intel / intel-gpu); put `-ngl all` in ENGINE_ARGS.
+# amd-gpu / intel-gpu VRAM quota is the gpu.bytetrade.io/*-gpu-memory pod
+# annotation (not nvidia.com/gpumem).
 apiVersion: apps/v1
 kind: Deployment
 metadata:
@@ -44,6 +64,7 @@ metadata:
     io.kompose.service: llamacpp
   annotations:
     applications.app.bytetrade.io/gpu-inject: "llamacpp"
+    llamacppllmbasev3.io/engine-image: {{ $engineImage | quote }}
 spec:
   replicas: {{ .Values.workloads.llamacppllmbasev3.replicaCount }}
   selector:
@@ -56,6 +77,13 @@ spec:
       labels:
         io.kompose.network/chrome-default: "true"
         io.kompose.service: llamacpp
+{{- if $needGpuMemAnnotation }}
+      # VRAM quota for the modes that cannot express it as a resource. Plain
+      # Kubernetes quantity (unlike nvidia.com/gpumem, which is a bare MiB int).
+      annotations:
+        gpu.bytetrade.io/required-gpu-memory: "{{ $gpuMiB }}Mi"
+        gpu.bytetrade.io/limited-gpu-memory: "{{ $gpuMiB }}Mi"
+{{- end }}
     spec:
       restartPolicy: Always
       volumes:
@@ -71,9 +99,35 @@ spec:
           configMap:
             name: llm-init-wrappers
             defaultMode: 0555
+{{- if $isAmdGpu }}
+        - name: kfd
+          hostPath:
+            path: /dev/kfd
+            type: CharDevice
+{{- end }}
+{{- if $needDri }}
+        - name: dri
+          hostPath:
+            path: /dev/dri
+            type: Directory
+        - name: drm
+          hostPath:
+            path: /sys/class/drm
+            type: Directory
+        - name: dshm
+          emptyDir:
+            medium: Memory
+            sizeLimit: {{ $shmLimit }}
+{{- end }}
+{{- if $isAmdGpu }}
+      securityContext:
+        supplementalGroups:
+          - 44
+          - 992
+{{- end }}
       containers:
         - name: llamacpp
-          image: docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10588
+          image: "{{ $engineImage }}"
           imagePullPolicy: IfNotPresent
           command: ["/bin/sh", "/llm-init/wrappers/llamacpp.sh"]
           env:
@@ -91,6 +145,8 @@ spec:
               value: "3600"
             - name: ENGINE_ARGS
               value: "{{ $engineArgs }}"
+            - name: OLARES_GPU_MODE
+              value: {{ $gpuType | quote }}
             - name: HF_HUB_CACHE
               value: /cache/hf/hub
 {{- if $hfEndpoint }}
@@ -125,22 +181,31 @@ spec:
             - name: wrappers
               mountPath: /llm-init/wrappers
               readOnly: true
+{{- if $isAmdGpu }}
+            - name: kfd
+              mountPath: /dev/kfd
+{{- end }}
+{{- if $needDri }}
+            - name: dri
+              mountPath: /dev/dri
+            - name: drm
+              mountPath: /sys/class/drm
+            - name: dshm
+              mountPath: /dev/shm
+{{- end }}
           resources:
-            # nvidia.com/gpumem is the per-instance GPU memory quota read by the
-            # Olares gpu-inject webhook (accelerator is -1/auto, can't carry it).
-            # It's a non-overcommittable extended resource, so it must be in BOTH
-            # requests and limits with the same value. Dropped on gb10 (unified
-            # memory uses pod memory instead) .
+            # nvidia.com/gpumem: nvidia only in chart; gb10 via pod memory +
+            # gpu-inject; amd-gpu / intel-gpu via the pod annotation above.
             requests:
               cpu: "{{ $cpuRequest }}"
               memory: "{{ $memRequest }}"
-{{- if not $unifiedMem }}
+{{- if $needNvidiaGpumem }}
               nvidia.com/gpumem: {{ $gpuMiB }}
 {{- end }}
             limits:
               cpu: "{{ $cpuLimit }}"
               memory: "{{ $memLimit }}"
-{{- if not $unifiedMem }}
+{{- if $needNvidiaGpumem }}
               nvidia.com/gpumem: {{ $gpuMiB }}
 {{- end }}
 ---
diff --git a/llamacppllmbasev3/templates/llm-init.yaml b/llamacppllmbasev3/templates/llm-init.yaml
index d069c644f..0f6165ddd 100644
--- a/llamacppllmbasev3/templates/llm-init.yaml
+++ b/llamacppllmbasev3/templates/llm-init.yaml
@@ -53,7 +53,7 @@
 {{- $logLevel := $oe.LOG_LEVEL | default "debug" -}}
 {{- $hfEndpoint := $oe.HF_ENDPOINT | default "" -}}
 {{- $hfToken := $oe.HF_TOKEN | default "" -}}
-{{- $runStateHostPath := printf "%s/huggingface/llm-init-run/%s" .Values.userspace.appCommon .Release.Name -}}
+{{- $runStateHostPath := printf "%s/llm-init-run/%s" .Values.userspace.appCache .Release.Name -}}
 ---
 # llm-init Pod (no GPU). Downloads GGUF models, writes sentinel + model_path,
 # serves progress UI + OpenAI /v1/* proxy on :8090, reverse-proxies the engine
@@ -109,7 +109,7 @@ spec:
           # v1.2.4 defaults to the stable LFS download path (hf_xet disabled
           # unless HF_ENABLE_XET=true); LFS streams each file to disk with a
           # bounded RAM footprint, avoiding the hf_xet OOM (huggingface_hub#3300).
-          image: docker.io/beclab/llm-init:v1.5.0
+          image: docker.io/beclab/llm-init:v1.7.12
           imagePullPolicy: IfNotPresent
           env:
             - name: ENGINE_KIND
diff --git a/llamacppllmbasev3/values.yaml b/llamacppllmbasev3/values.yaml
index c6133968b..ce9cbb804 100644
--- a/llamacppllmbasev3/values.yaml
+++ b/llamacppllmbasev3/values.yaml
@@ -7,3 +7,13 @@ workloads:
     replicaCount: 1
   llamacppllmbasev3:
     replicaCount: 1
+
+# llama.cpp engine images by Olares accelerator mode.
+engine:
+  images:
+    nvidia: docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10969
+    nvidiaArm64: docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10969
+    nvidiaGb10: docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10969
+    amdGpu: docker.io/beclab/ggml-org-llama.cpp:server-rocm-b10731
+    intel: docker.io/beclab/ggml-org-llama.cpp:server-intel-b10884
+    intelGpu: docker.io/beclab/ggml-org-llama.cpp:server-intel-b10752

From 6f12a4a10c5da52724e6d8f3fec6dc775482ec5f Mon Sep 17 00:00:00 2001
From: Harvey <harvey@bytetrade.io>
Date: Tue, 15 Sep 2026 12:10:10 +0800
Subject: [PATCH 2/3] Limit update to public baseline CUDA image bump

---
 llamacppllmbasev3/Chart.yaml                  |  2 +-
 llamacppllmbasev3/OlaresManifest.yaml         | 55 ++---------
 .../i18n/de-DE/OlaresManifest.yaml            | 40 --------
 .../i18n/en-US/OlaresManifest.yaml            |  7 +-
 .../i18n/es-ES/OlaresManifest.yaml            | 40 --------
 .../i18n/fr-FR/OlaresManifest.yaml            | 40 --------
 .../i18n/it-IT/OlaresManifest.yaml            | 40 --------
 .../i18n/ja-JP/OlaresManifest.yaml            | 40 --------
 .../i18n/zh-CN/OlaresManifest.yaml            |  7 +-
 llamacppllmbasev3/templates/_helpers.tpl      | 50 ----------
 llamacppllmbasev3/templates/llamacpp.yaml     | 97 +++----------------
 llamacppllmbasev3/templates/llm-init.yaml     |  4 +-
 llamacppllmbasev3/values.yaml                 | 10 --
 13 files changed, 31 insertions(+), 401 deletions(-)
 delete mode 100644 llamacppllmbasev3/i18n/de-DE/OlaresManifest.yaml
 delete mode 100644 llamacppllmbasev3/i18n/es-ES/OlaresManifest.yaml
 delete mode 100644 llamacppllmbasev3/i18n/fr-FR/OlaresManifest.yaml
 delete mode 100644 llamacppllmbasev3/i18n/it-IT/OlaresManifest.yaml
 delete mode 100644 llamacppllmbasev3/i18n/ja-JP/OlaresManifest.yaml

diff --git a/llamacppllmbasev3/Chart.yaml b/llamacppllmbasev3/Chart.yaml
index f9d32c363..53c28ebb3 100644
--- a/llamacppllmbasev3/Chart.yaml
+++ b/llamacppllmbasev3/Chart.yaml
@@ -3,4 +3,4 @@ appVersion: b10969
 description: Generic llama.cpp + llm-init base; the GGUF model is supplied at install time via env
 name: llamacppllmbasev3
 type: application
-version: 1.2.47
+version: 1.2.38
\ No newline at end of file
diff --git a/llamacppllmbasev3/OlaresManifest.yaml b/llamacppllmbasev3/OlaresManifest.yaml
index 3caa50aeb..99e79eb4e 100644
--- a/llamacppllmbasev3/OlaresManifest.yaml
+++ b/llamacppllmbasev3/OlaresManifest.yaml
@@ -7,7 +7,7 @@ metadata:
   description: "Generic llama.cpp engine base. Pick any GGUF model at install via env."
   appid: llamacppllmbasev3
   title: llama.cpp Engine Base
-  version: '1.2.47'
+  version: '1.2.38'
   categories:
     - AI
     - models
@@ -36,9 +36,9 @@ workloadReplicas:
   llminit: 1
 spec:
   onlyAdmin: true
-  versionName: 'b10969 / b10731-rocm / b10752-intel / b10884-intel-iGPU'
+  versionName: 'b10969'
   upgradeDescription: |
-    Chart 1.2.47 — Upgrade the NVIDIA / Spark CUDA image from server-cuda12-b10752 to server-cuda12-b10969.
+    Chart 1.2.38 — Upgrade the llama.cpp CUDA image from server-cuda12-b10588 to server-cuda12-b10969.
 
   fullDescription: |
     **IMPORTANT NOTE**  
@@ -51,7 +51,7 @@ spec:
     - `MODEL_SUPPORTS`: Model capability options (Vision / Tools / Thinking / None). Pick what your model supports. This is required.
     - `ENGINE_ARGS`: `-c 8192 -ngl all -fa on`. Extra llama.cpp options, like `-c 8192 -ngl all -fa on`. To fully use GPU, add `-ngl all`. You can also use values like `LLAMA_ARG_*=val`.
     - `LOG_LEVEL`: `debug`, `info`, `warn`, or `error`. Log level: choose from `debug`, `info`, `warn`, or `error`.
-    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192`, or `8192Mi`. Model memory needed. Accepts values like `8Gi`, `8192`, or `8192Mi`. On NVIDIA, this sets `nvidia.com/gpumem` in MiB; on Spark and `intel` integrated GPUs, sets pod memory (include model, KV cache and runtime overhead); on `amd-gpu` and `intel-gpu` declares the discrete VRAM quota through the `gpu.bytetrade.io/required-gpu-memory` pod annotation.
+    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192`, or `8192Mi`. Model memory needed. Accepts values like `8Gi`, `8192`, or `8192Mi`. On NVIDIA, this sets `nvidia.com/gpumem` in MiB; on Spark, sets pod memory.
     
     **How It Works**
     First, `llm-init` downloads your model and gets it ready, then `llama.cpp` starts serving the model when it’s ready. These two parts run separately but share files and progress info.
@@ -65,9 +65,6 @@ spec:
     |----------------|-------------------|---------------------------------------------------------------------|
     | `nvidia`       | GPU (CUDA)        | Set `LLAMACPP_REQUIRED_GPU_MEMORY`; use `-ngl all` for full GPU use |
     | `nvidia-gb10`  | GPU (Spark)       | Like above, but memory is managed by the pod                        |
-    | `amd-gpu`      | GPU (ROCm)        | Discrete AMD (e.g. Radeon AI PRO R9700 / gfx1201). Image `server-rocm-b10731`. Use `-ngl all`. Do not use `amd` (Ryzen iGPU). |
-    | `intel` | Intel integrated GPU (Panther Lake / SYCL) | `server-intel-b10884`; `-ngl all`; LLAMACPP_REQUIRED_GPU_MEMORY = shared pod memory (model + KV + runtime). |
-    | `intel-gpu`    | GPU (SYCL)        | Discrete Intel Arc (e.g. Arc Pro B70). Image `server-intel-b10752`. Use `-ngl all`. |
 
     **Model Storage**
     All model files are stored in a shared folder at `appCommon/huggingface`, mounted as `/cache/hf/hub`. All similar apps use this location for models.
@@ -83,23 +80,13 @@ spec:
   locale:
     - en-US
     - zh-CN
-    - de-DE
-    - es-ES
-    - it-IT
-    - fr-FR
-    - ja-JP
   license:
     - text: MIT
       url: https://github.com/ggml-org/llama.cpp?tab=MIT-1-ov-file
 
   accelerator:
-    #   nvidia       - discrete GPU; LLAMACPP_REQUIRED_GPU_MEMORY -> nvidia.com/gpumem
-    #   nvidia-gb10  - DGX Spark unified memory; same env -> pod memory
-    #   amd-gpu      - discrete AMD ROCm; LLAMACPP_REQUIRED_GPU_MEMORY -> gpu-memory pod annotation
-    #   intel-gpu    - discrete Intel Arc SYCL; LLAMACPP_REQUIRED_GPU_MEMORY -> gpu-memory pod annotation
-    # The two discrete non-NVIDIA modes have no GPU-memory extended resource to
-    # carry the quota, so the chart puts it on gpu.bytetrade.io/required-gpu-memory
-    # and app-service reads that to resolve requiredGPUMemory: "-1".
+    #   nvidia       — discrete GPU; LLAMACPP_REQUIRED_GPU_MEMORY -> gpumem
+    #   nvidia-gb10  — DGX Spark unified memory; same env -> pod memory
     - mode: nvidia
       limitedCpu: "-1"
       requiredCpu: "-1"
@@ -116,37 +103,11 @@ spec:
       limitedDisk: 500Gi
       limitedMemory: "-1"
       requiredMemory: "-1"
-    - mode: amd-gpu
-      limitedCpu: "-1"
-      requiredCpu: "-1"
-      requiredDisk: 50Mi
-      limitedDisk: 500Gi
-      limitedMemory: "-1"
-      requiredMemory: "-1"
-      requiredGPUMemory: "-1"
-      limitedGPUMemory: "-1"
-    - mode: intel
-      limitedCpu: "-1"
-      requiredCpu: "-1"
-      requiredDisk: 50Mi
-      limitedDisk: 500Gi
-      limitedMemory: "-1"
-      requiredMemory: "-1"
-    - mode: intel-gpu
-      limitedCpu: "-1"
-      requiredCpu: "-1"
-      requiredDisk: 50Mi
-      limitedDisk: 500Gi
-      limitedMemory: "-1"
-      requiredMemory: "-1"
-      requiredGPUMemory: "-1"
-      limitedGPUMemory: "-1"
 
   supportArch:
     - amd64
     - arm64
 permission:
-  appCache: true
   appData: true
   appCommon: true
 options:
@@ -209,7 +170,7 @@ envs:
     type: string
     editable: true
     applyOnChange: true
-    description:  "Example: -c 8192 -ngl all -fa on. Extra llama.cpp options, to fully use GPU, add `-ngl all`. You can also use values like `LLAMA_ARG_*=val`. Note that `-c` sizes the whole KV cache rather than one request, and omitting `-np` selects llama.cpp's default of 4 slots sharing it — set `-np 1` for a single long conversation, or raise `-c` to match the concurrency you want."
+    description:  "Example: -c 8192 -ngl all -fa on. Extra llama.cpp options, to fully use GPU, add `-ngl all`. You can also use values like `LLAMA_ARG_*=val`."
   - envName: LOG_LEVEL           # single-select: debug | info | warn | error
     required: false
     type: string
@@ -264,5 +225,5 @@ envs:
     type: string
     editable: false
     applyOnChange: false
-    description:  "Example: `8Gi`, `8192`, or `8192Mi`. GPU memory required for the model. On NVIDIA this sets `nvidia.com/gpumem` in MiB. On Spark and intel integrated GPUs it sets pod memory; include model, KV cache and runtime overhead. On amd-gpu and intel-gpu it declares the discrete VRAM quota through the gpu.bytetrade.io/required-gpu-memory pod annotation."
+    description:  "Example: `8Gi`, `8192`, or `8192Mi`. GPU memory required for the model. On NVIDIA this sets `nvidia.com/gpumem` in MiB. On Spark it sets pod memory."
     regex: "^[0-9]+(Gi|Mi)?$"
diff --git a/llamacppllmbasev3/i18n/de-DE/OlaresManifest.yaml b/llamacppllmbasev3/i18n/de-DE/OlaresManifest.yaml
deleted file mode 100644
index 796ca5739..000000000
--- a/llamacppllmbasev3/i18n/de-DE/OlaresManifest.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-metadata:
-  description: "Generische llama.cpp-Engine-Basis. Beliebiges GGUF-Modell bei der Installation per Env wählen"
-  title: llama.cpp Engine Base
-
-spec:
-  upgradeDescription: |
-    Chart 1.2.47 — NVIDIA-/Spark-CUDA-Image von server-cuda12-b10752 auf server-cuda12-b10969 aktualisiert.
-
-  fullDescription: |
-    **WICHTIGER HINWEIS**  
-    Diese App ist eine Vorlage und kann nicht allein genutzt werden. Setzen Sie die unten stehenden Umgebungsvariablen, um ein konkretes Modell anzubinden.
-
-    **Einrichtungsleitfaden für Umgebungsvariablen**
-    - `MODEL_SOURCE`: `hf://<repo> --include <file>.gguf`. Quelle zum Herunterladen Ihres Modells. Bei Split-Dateien nur den ersten Teil angeben; bei Vision-Language-Modellen den Projektor per Komma hinzufügen, z. B. `hf://<repo> --include model.gguf,hf://<repo> --include mmproj-F16.gguf`.
-    - `MODEL_NAME`: `owner/repo:quant`. Name Ihres Modells im Format `owner/repo:quant`, z. B. `unsloth/Qwen3.5-2B-GGUF:UD-Q4_K_XL`. Wird auch als OpenAI-Alias verwendet.
-    - `MODEL_MODE`: `chat` oder `embedding`. Bei embedding fügt das Chart automatisch `--embedding` hinzu.
-    - `MODEL_SUPPORTS`: Modellfähigkeiten (Vision / Tools / Thinking / None). Wählen Sie, was Ihr Modell unterstützt. Erforderlich.
-    - `ENGINE_ARGS`: `-c 8192 -ngl all -fa on`. Zusätzliche llama.cpp-Optionen wie `-c 8192 -ngl all -fa on`. Für volle GPU-Nutzung `-ngl all` hinzufügen. Auch Werte wie `LLAMA_ARG_*=val` möglich.
-    - `LOG_LEVEL`: `debug`, `info`, `warn` oder `error`. Log-Level: Wählen Sie zwischen `debug`, `info`, `warn` oder `error`.
-    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192` oder `8192Mi`. Benötigter Modellspeicher. Akzeptiert Werte wie `8Gi`, `8192` oder `8192Mi`. Auf NVIDIA setzt dies `nvidia.com/gpumem` in MiB; auf Spark die Pod-Memory; auf `amd-gpu` und `intel-gpu` deklariert es das diskrete VRAM-Kontingent über die Pod-Annotation `gpu.bytetrade.io/required-gpu-memory`.
-
-    **Funktionsweise**
-    Zuerst lädt `llm-init` Ihr Modell herunter und bereitet es vor, dann startet `llama.cpp` den Serving, wenn es bereit ist. Beide Teile laufen getrennt, teilen aber Dateien und Fortschrittsinformationen.
-
-    **Stack**
-    - **llama.cpp**: Führt das Modell aus und bedient API-Anfragen (Port 8081)
-    - **llm-init**: Lädt das Modell herunter, zeigt Fortschritt und stellt die OpenAI-API auf Port 8090 bereit, mit Proxy zu `llamacpp`.
-
-    **Beschleuniger-Modi**
-    | Modus          | Gerät             | Details                                                            |
-    |----------------|-------------------|---------------------------------------------------------------------|
-    | `nvidia`       | GPU (CUDA)        | `LLAMACPP_REQUIRED_GPU_MEMORY` setzen; `-ngl all` für volle GPU-Nutzung |
-    | `nvidia-gb10`  | GPU (Spark)       | Wie oben, aber Speicher wird vom Pod verwaltet                        |
-
-    **Modellspeicher**
-    Alle Modelldateien liegen im gemeinsamen Ordner `appCommon/huggingface`, gemountet als `/cache/hf/hub`. Ähnliche Apps nutzen denselben Ort für Modelle.
-
-    **Ressourcentipps**
-    - CPU, RAM und GPU konfigurieren sich automatisch anhand des Charts; Disk-Einstellungen sind fest.
-    - Für die GPU `-ngl all` in `ENGINE_ARGS` setzen. `-ngl all` für volles GPU-Offload.
diff --git a/llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml b/llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml
index e958e8d74..cf78b236b 100644
--- a/llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml
+++ b/llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml
@@ -4,7 +4,7 @@ metadata:
 
 spec:
   upgradeDescription: |
-    Chart 1.2.47 — Upgrade the NVIDIA / Spark CUDA image from server-cuda12-b10752 to server-cuda12-b10969.
+    Chart 1.2.38 — Upgrade the llama.cpp CUDA image from server-cuda12-b10588 to server-cuda12-b10969.
 
   fullDescription: |
     **IMPORTANT NOTE**  
@@ -17,7 +17,7 @@ spec:
     - `MODEL_SUPPORTS`: Model capability options (Vision / Tools / Thinking / None). Pick what your model supports. This is required.
     - `ENGINE_ARGS`: `-c 8192 -ngl all -fa on`. Extra llama.cpp options, like `-c 8192 -ngl all -fa on`. To fully use GPU, add `-ngl all`. You can also use values like `LLAMA_ARG_*=val`.
     - `LOG_LEVEL`: `debug`, `info`, `warn`, or `error`. Log level: choose from `debug`, `info`, `warn`, or `error`.
-    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192`, or `8192Mi`. Model memory needed. Accepts values like `8Gi`, `8192`, or `8192Mi`. On NVIDIA, this sets `nvidia.com/gpumem` in MiB; on Spark, sets pod memory; on `amd-gpu` and `intel-gpu` declares the discrete VRAM quota through the `gpu.bytetrade.io/required-gpu-memory` pod annotation.
+    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192`, or `8192Mi`. Model memory needed. Accepts values like `8Gi`, `8192`, or `8192Mi`. On NVIDIA, this sets `nvidia.com/gpumem` in MiB; on Spark, sets pod memory.
     
     **How It Works**
     First, `llm-init` downloads your model and gets it ready, then `llama.cpp` starts serving the model when it’s ready. These two parts run separately but share files and progress info.
@@ -31,9 +31,6 @@ spec:
     |----------------|-------------------|---------------------------------------------------------------------|
     | `nvidia`       | GPU (CUDA)        | Set `LLAMACPP_REQUIRED_GPU_MEMORY`; use `-ngl all` for full GPU use |
     | `nvidia-gb10`  | GPU (Spark)       | Like above, but memory is managed by the pod                        |
-    | `amd-gpu`      | GPU (ROCm)        | Discrete AMD (R9700 / gfx1201). Image `server-rocm-b10731`. Do not use `amd`. |
-    | `intel` | Intel integrated GPU (Panther Lake / SYCL) | `server-intel-b10884`; `-ngl all`; LLAMACPP_REQUIRED_GPU_MEMORY = shared pod memory (model + KV + runtime). |
-    | `intel-gpu`    | GPU (SYCL)        | Discrete Intel Arc (e.g. Arc Pro B70). Image `server-intel-b10752`. Use `-ngl all`. |
 
     **Model Storage**
     All model files are stored in a shared folder at `appCommon/huggingface`, mounted as `/cache/hf/hub`. All similar apps use this location for models.
diff --git a/llamacppllmbasev3/i18n/es-ES/OlaresManifest.yaml b/llamacppllmbasev3/i18n/es-ES/OlaresManifest.yaml
deleted file mode 100644
index a3f1b8340..000000000
--- a/llamacppllmbasev3/i18n/es-ES/OlaresManifest.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-metadata:
-  description: "Base genérica del motor llama.cpp. Elija cualquier modelo GGUF en la instalación vía env"
-  title: llama.cpp Engine Base
-
-spec:
-  upgradeDescription: |
-    Chart 1.2.47 — Actualiza la imagen CUDA de NVIDIA / Spark de server-cuda12-b10752 a server-cuda12-b10969.
-
-  fullDescription: |
-    **NOTA IMPORTANTE**  
-    Esta app es una plantilla y no puede usarse sola. Para usarla, configure las variables de entorno siguientes para conectar un modelo concreto.
-
-    **Guía de configuración de variables de entorno**
-    - `MODEL_SOURCE`: `hf://<repo> --include <file>.gguf`. Fuente para descargar su modelo. En archivos partidos, incluya solo la primera parte; en modelos visión-lenguaje, añada el proyector con una coma, p. ej. `hf://<repo> --include model.gguf,hf://<repo> --include mmproj-F16.gguf`.
-    - `MODEL_NAME`: `owner/repo:quant`. Nombre de su modelo en formato `owner/repo:quant`, p. ej. `unsloth/Qwen3.5-2B-GGUF:UD-Q4_K_XL`. También se usa como alias OpenAI.
-    - `MODEL_MODE`: `chat` o `embedding`. Si elige embedding, el chart añade `--embedding` automáticamente.
-    - `MODEL_SUPPORTS`: Capacidades del modelo (Vision / Tools / Thinking / None). Elija lo que su modelo admite. Obligatorio.
-    - `ENGINE_ARGS`: `-c 8192 -ngl all -fa on`. Opciones extra de llama.cpp, como `-c 8192 -ngl all -fa on`. Para usar la GPU al completo, añada `-ngl all`. También valores como `LLAMA_ARG_*=val`.
-    - `LOG_LEVEL`: `debug`, `info`, `warn` o `error`. Nivel de log: elija entre `debug`, `info`, `warn` o `error`.
-    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192` o `8192Mi`. Memoria de modelo necesaria. Admite valores como `8Gi`, `8192` o `8192Mi`. En NVIDIA establece `nvidia.com/gpumem` en MiB; en Spark, la memoria del pod; en `amd-gpu` e `intel-gpu` declara la cuota de VRAM discreta mediante la anotación de pod `gpu.bytetrade.io/required-gpu-memory`.
-
-    **Cómo funciona**
-    Primero, `llm-init` descarga su modelo y lo prepara; luego `llama.cpp` empieza a servir el modelo cuando está listo. Ambas partes se ejecutan por separado pero comparten archivos e información de progreso.
-
-    **Stack**
-    - **llama.cpp**: Ejecuta el modelo y atiende peticiones API (puerto 8081)
-    - **llm-init**: Descarga el modelo, muestra progreso y proporciona la API OpenAI en el puerto 8090, haciendo proxy a `llamacpp`.
-
-    **Modos de acelerador**
-    | Modo           | Dispositivo       | Detalles                                                           |
-    |----------------|-------------------|---------------------------------------------------------------------|
-    | `nvidia`       | GPU (CUDA)        | Establezca `LLAMACPP_REQUIRED_GPU_MEMORY`; use `-ngl all` para GPU completa |
-    | `nvidia-gb10`  | GPU (Spark)       | Como arriba, pero la memoria la gestiona el pod                       |
-
-    **Almacenamiento de modelos**
-    Todos los archivos de modelo se guardan en la carpeta compartida `appCommon/huggingface`, montada como `/cache/hf/hub`. Apps similares usan esta ubicación para modelos.
-
-    **Consejos de recursos**
-    - CPU, RAM y GPU se autoconfiguran según el chart; los ajustes de disco son fijos.
-    - Para usar la GPU, establezca `-ngl all` en `ENGINE_ARGS`. Use `-ngl all` para offload GPU completo.
diff --git a/llamacppllmbasev3/i18n/fr-FR/OlaresManifest.yaml b/llamacppllmbasev3/i18n/fr-FR/OlaresManifest.yaml
deleted file mode 100644
index 00d610e5b..000000000
--- a/llamacppllmbasev3/i18n/fr-FR/OlaresManifest.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-metadata:
-  description: "Base générique du moteur llama.cpp. Choisissez n'importe quel modèle GGUF à l'installation via env"
-  title: llama.cpp Engine Base
-
-spec:
-  upgradeDescription: |
-    Chart 1.2.47 — Mise à jour CUDA NVIDIA / Spark de server-cuda12-b10752 vers server-cuda12-b10969.
-
-  fullDescription: |
-    **NOTE IMPORTANTE**  
-    Cette app est un modèle et ne peut pas être utilisée seule. Pour l'utiliser, définissez les variables d'environnement ci-dessous pour connecter un modèle spécifique.
-
-    **Guide de configuration des variables d'environnement**
-    - `MODEL_SOURCE` : `hf://<repo> --include <file>.gguf`. Source pour télécharger votre modèle. Pour les fichiers fractionnés, n'incluez que la première partie ; pour les modèles vision-langage, ajoutez le projecteur avec une virgule, ex. `hf://<repo> --include model.gguf,hf://<repo> --include mmproj-F16.gguf`.
-    - `MODEL_NAME` : `owner/repo:quant`. Nom de votre modèle au format `owner/repo:quant`, ex. `unsloth/Qwen3.5-2B-GGUF:UD-Q4_K_XL`. Aussi utilisé comme alias OpenAI.
-    - `MODEL_MODE` : `chat` ou `embedding`. Si vous choisissez embedding, le chart ajoute automatiquement `--embedding`.
-    - `MODEL_SUPPORTS` : Capacités du modèle (Vision / Tools / Thinking / None). Choisissez ce que votre modèle prend en charge. Obligatoire.
-    - `ENGINE_ARGS` : `-c 8192 -ngl all -fa on`. Options llama.cpp supplémentaires, comme `-c 8192 -ngl all -fa on`. Pour utiliser pleinement le GPU, ajoutez `-ngl all`. Aussi des valeurs comme `LLAMA_ARG_*=val`.
-    - `LOG_LEVEL` : `debug`, `info`, `warn` ou `error`. Niveau de log : choisissez parmi `debug`, `info`, `warn` ou `error`.
-    - `LLAMACPP_REQUIRED_GPU_MEMORY` : `8Gi`, `8192` ou `8192Mi`. Mémoire modèle nécessaire. Accepte des valeurs comme `8Gi`, `8192` ou `8192Mi`. Sur NVIDIA, définit `nvidia.com/gpumem` en MiB ; sur Spark, la mémoire du pod ; sur `amd-gpu` et `intel-gpu`, déclare le quota de VRAM discrète via l'annotation de pod `gpu.bytetrade.io/required-gpu-memory`.
-
-    **Fonctionnement**
-    D'abord, `llm-init` télécharge votre modèle et le prépare, puis `llama.cpp` commence à servir le modèle quand il est prêt. Les deux parties tournent séparément mais partagent fichiers et infos de progression.
-
-    **Stack**
-    - **llama.cpp** : Exécute le modèle et sert les requêtes API (port 8081)
-    - **llm-init** : Télécharge le modèle, affiche la progression et fournit l'API OpenAI sur le port 8090, en proxy vers `llamacpp`.
-
-    **Modes accélérateur**
-    | Mode           | Périphérique      | Détails                                                            |
-    |----------------|-------------------|---------------------------------------------------------------------|
-    | `nvidia`       | GPU (CUDA)        | Définir `LLAMACPP_REQUIRED_GPU_MEMORY` ; utiliser `-ngl all` pour GPU complet |
-    | `nvidia-gb10`  | GPU (Spark)       | Comme ci-dessus, mais la mémoire est gérée par le pod                 |
-
-    **Stockage des modèles**
-    Tous les fichiers modèle sont stockés dans le dossier partagé `appCommon/huggingface`, monté comme `/cache/hf/hub`. Des apps similaires utilisent cet emplacement pour les modèles.
-
-    **Conseils ressources**
-    - CPU, RAM et GPU s'autoconfigurent selon le chart ; les réglages disque sont fixes.
-    - Pour utiliser le GPU, définissez `-ngl all` dans `ENGINE_ARGS`. Utilisez `-ngl all` pour un offload GPU complet.
diff --git a/llamacppllmbasev3/i18n/it-IT/OlaresManifest.yaml b/llamacppllmbasev3/i18n/it-IT/OlaresManifest.yaml
deleted file mode 100644
index c65b2eb9c..000000000
--- a/llamacppllmbasev3/i18n/it-IT/OlaresManifest.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-metadata:
-  description: "Base generica del motore llama.cpp. Scegli qualsiasi modello GGUF in installazione via env"
-  title: llama.cpp Engine Base
-
-spec:
-  upgradeDescription: |
-    Chart 1.2.47 — Aggiornamento CUDA NVIDIA / Spark da server-cuda12-b10752 a server-cuda12-b10969.
-
-  fullDescription: |
-    **NOTA IMPORTANTE**  
-    Questa app è un template e non può essere usata da sola. Per usarla, imposta le variabili d'ambiente sotto per collegare un modello specifico.
-
-    **Guida alla configurazione delle variabili d'ambiente**
-    - `MODEL_SOURCE`: `hf://<repo> --include <file>.gguf`. Sorgente da cui scaricare il modello. Per file suddivisi includi solo la prima parte; per modelli vision-language aggiungi il proiettore con una virgola, es. `hf://<repo> --include model.gguf,hf://<repo> --include mmproj-F16.gguf`.
-    - `MODEL_NAME`: `owner/repo:quant`. Nome del modello in formato `owner/repo:quant`, es. `unsloth/Qwen3.5-2B-GGUF:UD-Q4_K_XL`. Usato anche come alias OpenAI.
-    - `MODEL_MODE`: `chat` o `embedding`. Se scegli embedding, il chart aggiunge automaticamente `--embedding`.
-    - `MODEL_SUPPORTS`: Capacità del modello (Vision / Tools / Thinking / None). Scegli ciò che il modello supporta. Obbligatorio.
-    - `ENGINE_ARGS`: `-c 8192 -ngl all -fa on`. Opzioni extra llama.cpp, come `-c 8192 -ngl all -fa on`. Per usare appieno la GPU, aggiungi `-ngl all`. Anche valori come `LLAMA_ARG_*=val`.
-    - `LOG_LEVEL`: `debug`, `info`, `warn` o `error`. Livello di log: scegli tra `debug`, `info`, `warn` o `error`.
-    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`, `8192` o `8192Mi`. Memoria modello necessaria. Accetta valori come `8Gi`, `8192` o `8192Mi`. Su NVIDIA imposta `nvidia.com/gpumem` in MiB; su Spark, la memoria del pod; su `amd-gpu` e `intel-gpu` dichiara la quota di VRAM discreta tramite l'annotazione del pod `gpu.bytetrade.io/required-gpu-memory`.
-
-    **Come funziona**
-    Prima `llm-init` scarica il modello e lo prepara, poi `llama.cpp` inizia a servirlo quando è pronto. Le due parti girano separatemente ma condividono file e info di progresso.
-
-    **Stack**
-    - **llama.cpp**: Esegue il modello e serve le richieste API (porta 8081)
-    - **llm-init**: Scarica il modello, mostra il progresso e fornisce l'API OpenAI sulla porta 8090, facendo proxy verso `llamacpp`.
-
-    **Modalità acceleratore**
-    | Modalità       | Dispositivo       | Dettagli                                                           |
-    |----------------|-------------------|---------------------------------------------------------------------|
-    | `nvidia`       | GPU (CUDA)        | Imposta `LLAMACPP_REQUIRED_GPU_MEMORY`; usa `-ngl all` per GPU piena |
-    | `nvidia-gb10`  | GPU (Spark)       | Come sopra, ma la memoria è gestita dal pod                          |
-
-    **Archiviazione modelli**
-    Tutti i file modello sono nella cartella condivisa `appCommon/huggingface`, montata come `/cache/hf/hub`. App simili usano questa posizione per i modelli.
-
-    **Suggerimenti sulle risorse**
-    - CPU, RAM e GPU si autoconfigurano in base al chart; le impostazioni disco sono fisse.
-    - Per usare la GPU, imposta `-ngl all` in `ENGINE_ARGS`. Usa `-ngl all` per offload GPU completo.
diff --git a/llamacppllmbasev3/i18n/ja-JP/OlaresManifest.yaml b/llamacppllmbasev3/i18n/ja-JP/OlaresManifest.yaml
deleted file mode 100644
index c70f4d8a4..000000000
--- a/llamacppllmbasev3/i18n/ja-JP/OlaresManifest.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-metadata:
-  description: "汎用 llama.cpp エンジンベース。インストール時に env で任意の GGUF モデルを選択"
-  title: llama.cpp Engine Base
-
-spec:
-  upgradeDescription: |
-    Chart 1.2.47 — NVIDIA / Spark CUDA イメージを server-cuda12-b10752 から server-cuda12-b10969 に更新。
-
-  fullDescription: |
-    **重要なお知らせ**  
-    このアプリはテンプレートであり、単体では使えません。下の環境変数を設定して特定モデルを接続してください。
-
-    **環境変数セットアップガイド**
-    - `MODEL_SOURCE`: `hf://<repo> --include <file>.gguf`。モデルのダウンロード元。分割ファイルは先頭パートのみ; ビジョン言語モデルはカンマでプロジェクタを追加。例: `hf://<repo> --include model.gguf,hf://<repo> --include mmproj-F16.gguf`。
-    - `MODEL_NAME`: `owner/repo:quant`。`owner/repo:quant` 形式のモデル名。例: `unsloth/Qwen3.5-2B-GGUF:UD-Q4_K_XL`。OpenAI エイリアスとしても使用。
-    - `MODEL_MODE`: `chat` または `embedding`。embedding を選ぶとチャートが自動で `--embedding` を追加。
-    - `MODEL_SUPPORTS`: モデル能力（Vision / Tools / Thinking / None）。モデルが対応するものを選択。必須。
-    - `ENGINE_ARGS`: `-c 8192 -ngl all -fa on`。`-c 8192 -ngl all -fa on` などの追加 llama.cpp オプション。GPU をフルに使うには `-ngl all`。`LLAMA_ARG_*=val` 形式も可。
-    - `LOG_LEVEL`: `debug`、`info`、`warn`、または `error`。ログレベルは `debug`、`info`、`warn`、`error` から選択。
-    - `LLAMACPP_REQUIRED_GPU_MEMORY`: `8Gi`、`8192`、または `8192Mi`。必要なモデルメモリ。`8Gi`、`8192`、`8192Mi` などの値を使用可能。NVIDIA では MiB 単位の `nvidia.com/gpumem` を設定; Spark ではポッドメモリを設定; `amd-gpu` と `intel-gpu` では Pod アノテーション `gpu.bytetrade.io/required-gpu-memory` でディスクリート VRAM クォータを宣言。
-
-    **動作の仕組み**
-    まず `llm-init` がモデルをダウンロードして準備し、準備ができたら `llama.cpp` がサービングを開始します。両者は別々に動作しますが、ファイルと進捗情報を共有します。
-
-    **スタック**
-    - **llama.cpp**: モデルを実行し API リクエストを処理（ポート 8081）
-    - **llm-init**: モデルをダウンロードし、進捗を表示し、ポート 8090 で OpenAI API を提供して `llamacpp` へプロキシ。
-
-    **アクセラレータモード**
-    | モード         | デバイス          | 詳細                                                               |
-    |----------------|-------------------|---------------------------------------------------------------------|
-    | `nvidia`       | GPU (CUDA)        | `LLAMACPP_REQUIRED_GPU_MEMORY` を設定; フル GPU には `-ngl all`     |
-    | `nvidia-gb10`  | GPU (Spark)       | 上記と同様だがメモリはポッドが管理                                    |
-
-    **モデルストレージ**
-    すべてのモデルファイルは共有フォルダ `appCommon/huggingface` に保存され、`/cache/hf/hub` にマウントされます。類似アプリもモデルにこの場所を使います。
-
-    **リソースのヒント**
-    - CPU、RAM、GPU はチャートに基づき自動設定; ディスク設定は固定。
-    - GPU を使うには `ENGINE_ARGS` に `-ngl all` を設定。フル GPU オフロードには `-ngl all`。
diff --git a/llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml b/llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml
index 85b5dba5e..752f2bb82 100644
--- a/llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml
+++ b/llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml
@@ -4,7 +4,7 @@ metadata:
 
 spec:
   upgradeDescription: |
-    Chart 1.2.47 — NVIDIA / Spark CUDA 镜像从 server-cuda12-b10752 升级到 server-cuda12-b10969。
+    Chart 1.2.38 — llama.cpp CUDA 镜像从 server-cuda12-b10588 升级到 server-cuda12-b10969。
 
   fullDescription: |
     **重要提示**  
@@ -17,7 +17,7 @@ spec:
     - `MODEL_SUPPORTS`：模型能力选项（Vision / Tools / Thinking / None）。请选择您的模型支持的功能，必填。
     - `ENGINE_ARGS`：`-c 8192 -ngl all -fa on`。如需完全利用 GPU，请加 `-ngl all`。也可用如 `LLAMA_ARG_*=val` 的参数。
     - `LOG_LEVEL`：`debug`、`info`、`warn` 或 `error`。日志等级，请从上述选项中选择。
-    - `LLAMACPP_REQUIRED_GPU_MEMORY`：`8Gi`、`8192` 或 `8192Mi`。模型（仅 GPU 模式）所需显存。支持 `8Gi`、`8192`、`8192Mi` 等格式；NVIDIA 下设置 `nvidia.com/gpumem`（MiB），Spark 下设置 pod 显存，`amd-gpu` 与 `intel-gpu` 下通过 pod annotation `gpu.bytetrade.io/required-gpu-memory` 声明独立显存配额。
+    - `LLAMACPP_REQUIRED_GPU_MEMORY`：`8Gi`、`8192` 或 `8192Mi`。模型（仅 GPU 模式）所需显存。支持 `8Gi`、`8192`、`8192Mi` 等格式；NVIDIA 下设置 `nvidia.com/gpumem`（MiB），Spark 下设置 pod 显存。
 
     **工作流程说明**
     首先，`llm-init` 下载并准备模型，然后 `llama.cpp` 在模型准备好后启动服务。这两个部分独立运行但共享文件与进度信息。
@@ -31,9 +31,6 @@ spec:
     |----------------|------------------|--------------------------------------------------------------|
     | `nvidia`       | GPU (CUDA)       | 需设置 `LLAMACPP_REQUIRED_GPU_MEMORY`，用 `-ngl all` 完全利用 GPU |
     | `nvidia-gb10`  | GPU (Spark)      | 类似上方，但显存由 pod 管理                                  |
-    | `amd-gpu`      | GPU (ROCm)       | AMD 独显（R9700 / gfx1201）。镜像 `server-rocm-b10731`。不要用 `amd`（核显）。 |
-    | `intel` | Intel integrated GPU (Panther Lake / SYCL) | `server-intel-b10884`; `-ngl all`; LLAMACPP_REQUIRED_GPU_MEMORY = shared pod memory (model + KV + runtime). |
-    | `intel-gpu`    | GPU (SYCL)       | Intel Arc 独显（如 Arc Pro B70）。镜像 `server-intel-b10752`。请用 `-ngl all`。 |
 
     **模型存储**
     所有模型文件存储在共享目录 `appCommon/huggingface`（挂载为 `/cache/hf/hub`），同类应用均使用此目录。
diff --git a/llamacppllmbasev3/templates/_helpers.tpl b/llamacppllmbasev3/templates/_helpers.tpl
index 71f04826f..3dcee99c8 100644
--- a/llamacppllmbasev3/templates/_helpers.tpl
+++ b/llamacppllmbasev3/templates/_helpers.tpl
@@ -23,53 +23,3 @@
 {{- $in := . -}}
 {{- trim ($in.Args | default "") -}}
 {{- end -}}
-{{- /* Olares GPU mode at install: nvidia | nvidia-gb10 | amd-gpu | intel | intel-gpu. */ -}}
-{{- define "llmbase.gpuType" -}}
-{{- $gpuObj := .Values.GPU | default dict -}}
-{{- $gpuType := .Values.gpu | default "" -}}
-{{- if not $gpuType -}}
-{{- $gpuType = $gpuObj.Type | default "nvidia" -}}
-{{- end -}}
-{{- $gpuType -}}
-{{- end -}}
-{{- /* llama.cpp engine image by accelerator. NVIDIA/Spark CUDA b10752; amd-gpu ROCm b10731; intel-gpu SYCL b10752. */ -}}
-{{- define "llamacppllmbasev3.engineImage" -}}
-{{- $gpuType := include "llmbase.gpuType" . -}}
-{{- $isGb10 := or (eq $gpuType "nvidia-gb10") (eq (include "llmbase.isGb10" .) "true") -}}
-{{- $img := .Values.engine.images | default dict -}}
-{{- if eq $gpuType "amd-gpu" -}}
-{{- $img.amdGpu | default "docker.io/beclab/ggml-org-llama.cpp:server-rocm-b10731" -}}
-{{- else if eq $gpuType "intel-gpu" -}}
-{{- $img.intelGpu | default "docker.io/beclab/ggml-org-llama.cpp:server-intel-b10752" -}}
-{{- else if eq $gpuType "intel" -}}
-{{- $img.intel | default "docker.io/beclab/ggml-org-llama.cpp:server-intel-b10884" -}}
-{{- else if $isGb10 -}}
-{{- $img.nvidiaGb10 | default "docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10969" -}}
-{{- else -}}
-{{- $img.nvidia | default "docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10969" -}}
-{{- end -}}
-{{- end -}}
-{{- /* Spark/GB10: detect from GPU.Type or node hardware (install-time .Values.nodes). */ -}}
-{{- define "llmbase.isGb10" -}}
-{{- $isGb10 := "false" -}}
-{{- if .Values.nodes -}}
-  {{- range $nodeIndex, $node := .Values.nodes -}}
-    {{- if eq $nodeIndex 0 -}}
-      {{- with $node -}}
-        {{- if .GPUS -}}
-          {{- range $gpuIndex, $gpu := .GPUS -}}
-            {{- if eq $gpuIndex 0 -}}
-              {{- with $gpu -}}
-                {{- if eq (upper .Model) "GB10" -}}
-                  {{- $isGb10 = "true" -}}
-                {{- end -}}
-              {{- end -}}
-            {{- end -}}
-          {{- end -}}
-        {{- end -}}
-      {{- end -}}
-    {{- end -}}
-  {{- end -}}
-{{- end -}}
-{{- $isGb10 -}}
-{{- end -}}
diff --git a/llamacppllmbasev3/templates/llamacpp.yaml b/llamacppllmbasev3/templates/llamacpp.yaml
index 141bf3f0a..f530357ca 100644
--- a/llamacppllmbasev3/templates/llamacpp.yaml
+++ b/llamacppllmbasev3/templates/llamacpp.yaml
@@ -13,46 +13,26 @@
 {{- $memRequest := trim ($oe.LLAMACPP_MEMORY_REQUEST | default "2Gi") -}}
 {{- $cpuLimit := trim ($oe.LLAMACPP_CPU_LIMIT | default "6") -}}
 {{- $memLimit := trim ($oe.LLAMACPP_MEMORY_LIMIT | default "35Gi") -}}
-{{- $gpuType := include "llmbase.gpuType" . -}}
-{{- $engineImage := include "llamacppllmbasev3.engineImage" . -}}
-{{- $gb10FromNodes := eq (include "llmbase.isGb10" .) "true" -}}
-{{- $isAmdGpu := eq $gpuType "amd-gpu" -}}
-{{- $isIntelGpu := eq $gpuType "intel-gpu" -}}
-{{- $isIntel := eq $gpuType "intel" -}}
-{{- $needDri := or $isAmdGpu $isIntelGpu $isIntel -}}
-{{- $unifiedMem := or $gb10FromNodes (eq $gpuType "nvidia-gb10") $isIntel -}}
-{{- /* nvidia.com/gpumem is NVIDIA-HAMi only: HAMi's extender exempts it from
-       the node capacity check, so it can carry a quota the kubelet never
-       advertises. The discrete AMD/Intel plugins advertise a card count and
-       nothing else, so their quota cannot ride on a container resource at all
-       and travels on a pod annotation instead, which is inert to the
-       scheduler. app-service reads it to resolve the requiredGPUMemory /
-       limitedGPUMemory "-1" sentinel this manifest declares for those modes;
-       without it the sentinel resolves to 0 and the card is dropped from its
-       own compute binding on resume. */ -}}
-{{- $needNvidiaGpumem := and (not $unifiedMem) (not $isAmdGpu) (not $isIntelGpu) -}}
-{{- $needGpuMemAnnotation := or $isAmdGpu $isIntelGpu -}}
-{{- $shmLimit := "16Gi" -}}
-{{- if or $isIntelGpu $isIntel -}}
-{{- $shmLimit = "10Gi" -}}
+{{- $gpuType := .Values.gpu | default "" -}}
+{{- if not $gpuType -}}
+{{- $gpuType = .Values.GPU.Type | default "nvidia" -}}
 {{- end -}}
+{{- $unifiedMem := eq $gpuType "nvidia-gb10" -}}
 {{- $engineArgs = include "llamacppllmbasev3.engineArgs" (dict "Args" $engineArgs) -}}
-{{- /* nvidia.com/gpumem needs a BARE MiB int; Spark / Intel integrated unified memory reuses the
+{{- /* nvidia.com/gpumem needs a BARE MiB int; Spark unified memory reuses the
        same value as pod memory ("<MiB>Mi"). */ -}}
 {{- $gpuMiB := include "llmbase.gpuMiB" ($oe.LLAMACPP_REQUIRED_GPU_MEMORY | default "4096") -}}
 {{- if $unifiedMem -}}
 {{- $memRequest = printf "%sMi" $gpuMiB -}}
 {{- $memLimit = printf "%sMi" $gpuMiB -}}
 {{- end -}}
-{{- $runStateHostPath := printf "%s/llm-init-run/%s" .Values.userspace.appCache .Release.Name -}}
+{{- $runStateHostPath := printf "%s/huggingface/llm-init-run/%s" .Values.userspace.appCommon .Release.Name -}}
 ---
 # llama.cpp engine Deployment. wrappers/llamacpp.sh blocks on the sentinel
 # (download ready), then execs llama-server -hf "$MODEL_NAME".
 # llm-init reverse-proxies at http://llamacpp:8081 — Service name MUST be
-# "llamacpp". GPU base: CUDA (nvidia / nvidia-gb10), ROCm (amd-gpu), or
-# Intel SYCL (intel / intel-gpu); put `-ngl all` in ENGINE_ARGS.
-# amd-gpu / intel-gpu VRAM quota is the gpu.bytetrade.io/*-gpu-memory pod
-# annotation (not nvidia.com/gpumem).
+# "llamacpp". GPU base: CUDA server build + gpu-inject; put `-ngl all` in
+# ENGINE_ARGS to offload all layers to the GPU.
 apiVersion: apps/v1
 kind: Deployment
 metadata:
@@ -64,7 +44,6 @@ metadata:
     io.kompose.service: llamacpp
   annotations:
     applications.app.bytetrade.io/gpu-inject: "llamacpp"
-    llamacppllmbasev3.io/engine-image: {{ $engineImage | quote }}
 spec:
   replicas: {{ .Values.workloads.llamacppllmbasev3.replicaCount }}
   selector:
@@ -77,13 +56,6 @@ spec:
       labels:
         io.kompose.network/chrome-default: "true"
         io.kompose.service: llamacpp
-{{- if $needGpuMemAnnotation }}
-      # VRAM quota for the modes that cannot express it as a resource. Plain
-      # Kubernetes quantity (unlike nvidia.com/gpumem, which is a bare MiB int).
-      annotations:
-        gpu.bytetrade.io/required-gpu-memory: "{{ $gpuMiB }}Mi"
-        gpu.bytetrade.io/limited-gpu-memory: "{{ $gpuMiB }}Mi"
-{{- end }}
     spec:
       restartPolicy: Always
       volumes:
@@ -99,35 +71,9 @@ spec:
           configMap:
             name: llm-init-wrappers
             defaultMode: 0555
-{{- if $isAmdGpu }}
-        - name: kfd
-          hostPath:
-            path: /dev/kfd
-            type: CharDevice
-{{- end }}
-{{- if $needDri }}
-        - name: dri
-          hostPath:
-            path: /dev/dri
-            type: Directory
-        - name: drm
-          hostPath:
-            path: /sys/class/drm
-            type: Directory
-        - name: dshm
-          emptyDir:
-            medium: Memory
-            sizeLimit: {{ $shmLimit }}
-{{- end }}
-{{- if $isAmdGpu }}
-      securityContext:
-        supplementalGroups:
-          - 44
-          - 992
-{{- end }}
       containers:
         - name: llamacpp
-          image: "{{ $engineImage }}"
+          image: docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10969
           imagePullPolicy: IfNotPresent
           command: ["/bin/sh", "/llm-init/wrappers/llamacpp.sh"]
           env:
@@ -145,8 +91,6 @@ spec:
               value: "3600"
             - name: ENGINE_ARGS
               value: "{{ $engineArgs }}"
-            - name: OLARES_GPU_MODE
-              value: {{ $gpuType | quote }}
             - name: HF_HUB_CACHE
               value: /cache/hf/hub
 {{- if $hfEndpoint }}
@@ -181,31 +125,22 @@ spec:
             - name: wrappers
               mountPath: /llm-init/wrappers
               readOnly: true
-{{- if $isAmdGpu }}
-            - name: kfd
-              mountPath: /dev/kfd
-{{- end }}
-{{- if $needDri }}
-            - name: dri
-              mountPath: /dev/dri
-            - name: drm
-              mountPath: /sys/class/drm
-            - name: dshm
-              mountPath: /dev/shm
-{{- end }}
           resources:
-            # nvidia.com/gpumem: nvidia only in chart; gb10 via pod memory +
-            # gpu-inject; amd-gpu / intel-gpu via the pod annotation above.
+            # nvidia.com/gpumem is the per-instance GPU memory quota read by the
+            # Olares gpu-inject webhook (accelerator is -1/auto, can't carry it).
+            # It's a non-overcommittable extended resource, so it must be in BOTH
+            # requests and limits with the same value. Dropped on gb10 (unified
+            # memory uses pod memory instead) .
             requests:
               cpu: "{{ $cpuRequest }}"
               memory: "{{ $memRequest }}"
-{{- if $needNvidiaGpumem }}
+{{- if not $unifiedMem }}
               nvidia.com/gpumem: {{ $gpuMiB }}
 {{- end }}
             limits:
               cpu: "{{ $cpuLimit }}"
               memory: "{{ $memLimit }}"
-{{- if $needNvidiaGpumem }}
+{{- if not $unifiedMem }}
               nvidia.com/gpumem: {{ $gpuMiB }}
 {{- end }}
 ---
diff --git a/llamacppllmbasev3/templates/llm-init.yaml b/llamacppllmbasev3/templates/llm-init.yaml
index 0f6165ddd..d069c644f 100644
--- a/llamacppllmbasev3/templates/llm-init.yaml
+++ b/llamacppllmbasev3/templates/llm-init.yaml
@@ -53,7 +53,7 @@
 {{- $logLevel := $oe.LOG_LEVEL | default "debug" -}}
 {{- $hfEndpoint := $oe.HF_ENDPOINT | default "" -}}
 {{- $hfToken := $oe.HF_TOKEN | default "" -}}
-{{- $runStateHostPath := printf "%s/llm-init-run/%s" .Values.userspace.appCache .Release.Name -}}
+{{- $runStateHostPath := printf "%s/huggingface/llm-init-run/%s" .Values.userspace.appCommon .Release.Name -}}
 ---
 # llm-init Pod (no GPU). Downloads GGUF models, writes sentinel + model_path,
 # serves progress UI + OpenAI /v1/* proxy on :8090, reverse-proxies the engine
@@ -109,7 +109,7 @@ spec:
           # v1.2.4 defaults to the stable LFS download path (hf_xet disabled
           # unless HF_ENABLE_XET=true); LFS streams each file to disk with a
           # bounded RAM footprint, avoiding the hf_xet OOM (huggingface_hub#3300).
-          image: docker.io/beclab/llm-init:v1.7.12
+          image: docker.io/beclab/llm-init:v1.5.0
           imagePullPolicy: IfNotPresent
           env:
             - name: ENGINE_KIND
diff --git a/llamacppllmbasev3/values.yaml b/llamacppllmbasev3/values.yaml
index ce9cbb804..c6133968b 100644
--- a/llamacppllmbasev3/values.yaml
+++ b/llamacppllmbasev3/values.yaml
@@ -7,13 +7,3 @@ workloads:
     replicaCount: 1
   llamacppllmbasev3:
     replicaCount: 1
-
-# llama.cpp engine images by Olares accelerator mode.
-engine:
-  images:
-    nvidia: docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10969
-    nvidiaArm64: docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10969
-    nvidiaGb10: docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10969
-    amdGpu: docker.io/beclab/ggml-org-llama.cpp:server-rocm-b10731
-    intel: docker.io/beclab/ggml-org-llama.cpp:server-intel-b10884
-    intelGpu: docker.io/beclab/ggml-org-llama.cpp:server-intel-b10752

From 47742574b955363b59328e73b93ba431a042421e Mon Sep 17 00:00:00 2001
From: Harvey <harvey@bytetrade.io>
Date: Tue, 15 Sep 2026 12:13:36 +0800
Subject: [PATCH 3/3] Document upstream b10588 to b10969 engine changes

---
 llamacppllmbasev3/OlaresManifest.yaml         | 22 ++++++++++++++++++-
 .../i18n/en-US/OlaresManifest.yaml            | 22 ++++++++++++++++++-
 .../i18n/zh-CN/OlaresManifest.yaml            | 22 ++++++++++++++++++-
 3 files changed, 63 insertions(+), 3 deletions(-)

diff --git a/llamacppllmbasev3/OlaresManifest.yaml b/llamacppllmbasev3/OlaresManifest.yaml
index 99e79eb4e..2e43bef8e 100644
--- a/llamacppllmbasev3/OlaresManifest.yaml
+++ b/llamacppllmbasev3/OlaresManifest.yaml
@@ -38,7 +38,27 @@ spec:
   onlyAdmin: true
   versionName: 'b10969'
   upgradeDescription: |
-    Chart 1.2.38 — Upgrade the llama.cpp CUDA image from server-cuda12-b10588 to server-cuda12-b10969.
+    Chart 1.2.38 — Upgrade llama.cpp from `server-cuda12-b10588` to `server-cuda12-b10969`.
+
+    **Model compatibility**
+    - Add Qwen3.8-Flash-Next (`qwen4exp`) text/vision architecture support ([#27742](https://github.com/ggml-org/llama.cpp/pull/27742)), with subsequent recurrent-state rollback and inference fixes.
+    - Add NVIDIA Nemotron-3-Puzzle-75B-A9B ([#25444](https://github.com/ggml-org/llama.cpp/pull/25444)) and Tencent Hy 4 preview architecture support ([#28127](https://github.com/ggml-org/llama.cpp/pull/28127)). These require matching model weights and sufficient resources.
+
+    **Speculative decoding and CUDA**
+    - Add DFlash2 draft-model support ([#27342](https://github.com/ggml-org/llama.cpp/pull/27342)); requires a compatible draft checkpoint. Upstream notes that DFlash2 GGUF files converted before August 27, 2026 need reconversion for vision correctness.
+    - Add CUDA sparse Flash Attention for DeepSeek-V4/GLM ([#27970](https://github.com/ggml-org/llama.cpp/pull/27970)), extend MoE fusion to speculative decoding ([#27621](https://github.com/ggml-org/llama.cpp/pull/27621)), and optimize Q4_K/Q5_K decoding with DGX Spark prefetch improvements ([#26705](https://github.com/ggml-org/llama.cpp/pull/26705)). Gains depend on model and hardware.
+
+    **Server and chat behavior**
+    - Add `--kv-unified-per-slot` to cap each request slot when using a unified KV cache; `-c` controls the total pool ([#24124](https://github.com/ggml-org/llama.cpp/pull/24124)). This update does not change the configured context size automatically.
+    - Enable `preserve_reasoning` by default for chat templates that support it; use `--no-reasoning-preserve` to disable ([#28174](https://github.com/ggml-org/llama.cpp/pull/28174)).
+    - Accept `data:` URLs in `input_audio` and `input_video` requests, fixing media decoding errors ([#27735](https://github.com/ggml-org/llama.cpp/pull/27735)); requires a compatible multimodal model and client.
+
+    **Stability and memory**
+    - Reduce peak RAM during model loading when tensor repacking is enabled ([#27483](https://github.com/ggml-org/llama.cpp/pull/27483)).
+    - Fix the model-router LRU hang with concurrent requests ([#28539](https://github.com/ggml-org/llama.cpp/pull/28539)), speculative decoding after image input ([#28715](https://github.com/ggml-org/llama.cpp/pull/28715)), and CUDA Flash Attention synchronization ([#27870](https://github.com/ggml-org/llama.cpp/pull/27870)).
+
+    These are upstream engine changes; model-specific capabilities are not all validated by this chart update. llm-init remains at `v1.5.0`.
+    Full upstream changes: https://github.com/ggml-org/llama.cpp/compare/b10588...b10969
 
   fullDescription: |
     **IMPORTANT NOTE**  
diff --git a/llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml b/llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml
index cf78b236b..6dac468a5 100644
--- a/llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml
+++ b/llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml
@@ -4,7 +4,27 @@ metadata:
 
 spec:
   upgradeDescription: |
-    Chart 1.2.38 — Upgrade the llama.cpp CUDA image from server-cuda12-b10588 to server-cuda12-b10969.
+    Chart 1.2.38 — Upgrade llama.cpp from `server-cuda12-b10588` to `server-cuda12-b10969`.
+
+    **Model compatibility**
+    - Add Qwen3.8-Flash-Next (`qwen4exp`) text/vision architecture support ([#27742](https://github.com/ggml-org/llama.cpp/pull/27742)), with subsequent recurrent-state rollback and inference fixes.
+    - Add NVIDIA Nemotron-3-Puzzle-75B-A9B ([#25444](https://github.com/ggml-org/llama.cpp/pull/25444)) and Tencent Hy 4 preview architecture support ([#28127](https://github.com/ggml-org/llama.cpp/pull/28127)). These require matching model weights and sufficient resources.
+
+    **Speculative decoding and CUDA**
+    - Add DFlash2 draft-model support ([#27342](https://github.com/ggml-org/llama.cpp/pull/27342)); requires a compatible draft checkpoint. Upstream notes that DFlash2 GGUF files converted before August 27, 2026 need reconversion for vision correctness.
+    - Add CUDA sparse Flash Attention for DeepSeek-V4/GLM ([#27970](https://github.com/ggml-org/llama.cpp/pull/27970)), extend MoE fusion to speculative decoding ([#27621](https://github.com/ggml-org/llama.cpp/pull/27621)), and optimize Q4_K/Q5_K decoding with DGX Spark prefetch improvements ([#26705](https://github.com/ggml-org/llama.cpp/pull/26705)). Gains depend on model and hardware.
+
+    **Server and chat behavior**
+    - Add `--kv-unified-per-slot` to cap each request slot when using a unified KV cache; `-c` controls the total pool ([#24124](https://github.com/ggml-org/llama.cpp/pull/24124)). This update does not change the configured context size automatically.
+    - Enable `preserve_reasoning` by default for chat templates that support it; use `--no-reasoning-preserve` to disable ([#28174](https://github.com/ggml-org/llama.cpp/pull/28174)).
+    - Accept `data:` URLs in `input_audio` and `input_video` requests, fixing media decoding errors ([#27735](https://github.com/ggml-org/llama.cpp/pull/27735)); requires a compatible multimodal model and client.
+
+    **Stability and memory**
+    - Reduce peak RAM during model loading when tensor repacking is enabled ([#27483](https://github.com/ggml-org/llama.cpp/pull/27483)).
+    - Fix the model-router LRU hang with concurrent requests ([#28539](https://github.com/ggml-org/llama.cpp/pull/28539)), speculative decoding after image input ([#28715](https://github.com/ggml-org/llama.cpp/pull/28715)), and CUDA Flash Attention synchronization ([#27870](https://github.com/ggml-org/llama.cpp/pull/27870)).
+
+    These are upstream engine changes; model-specific capabilities are not all validated by this chart update. llm-init remains at `v1.5.0`.
+    Full upstream changes: https://github.com/ggml-org/llama.cpp/compare/b10588...b10969
 
   fullDescription: |
     **IMPORTANT NOTE**  
diff --git a/llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml b/llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml
index 752f2bb82..b6eaeaab1 100644
--- a/llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml
+++ b/llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml
@@ -4,7 +4,27 @@ metadata:
 
 spec:
   upgradeDescription: |
-    Chart 1.2.38 — llama.cpp CUDA 镜像从 server-cuda12-b10588 升级到 server-cuda12-b10969。
+    Chart 1.2.38 — llama.cpp 镜像从 `server-cuda12-b10588` 升级到 `server-cuda12-b10969`。
+
+    **模型兼容性**
+    - 新增 Qwen3.8-Flash-Next（`qwen4exp`）文本与视觉架构支持（[#27742](https://github.com/ggml-org/llama.cpp/pull/27742)），并包含后续循环状态回滚和推理修复。
+    - 新增 NVIDIA Nemotron-3-Puzzle-75B-A9B（[#25444](https://github.com/ggml-org/llama.cpp/pull/25444)）和 Tencent Hy 4 预览架构支持（[#28127](https://github.com/ggml-org/llama.cpp/pull/28127)）；使用时需准备匹配的模型权重和足够的资源。
+
+    **推测解码与 CUDA**
+    - 新增 DFlash2 草稿模型支持（[#27342](https://github.com/ggml-org/llama.cpp/pull/27342)），需配置兼容的草稿权重。上游提示：2026 年 8 月 27 日前转换的 DFlash2 GGUF 文件需重新转换，以保证视觉功能正确。
+    - 新增 DeepSeek-V4/GLM 的 CUDA 稀疏 Flash Attention（[#27970](https://github.com/ggml-org/llama.cpp/pull/27970)），将 MoE 融合优化扩展到推测解码（[#27621](https://github.com/ggml-org/llama.cpp/pull/27621)），优化 Q4_K/Q5_K 解码和 DGX Spark 预取（[#26705](https://github.com/ggml-org/llama.cpp/pull/26705)）。实际收益取决于模型与硬件。
+
+    **服务接口与对话行为**
+    - 新增 `--kv-unified-per-slot`，可在统一 KV 缓存模式下限制每个请求槽的上下文，`-c` 控制总缓存池（[#24124](https://github.com/ggml-org/llama.cpp/pull/24124)）。本次升级不会自动修改实例的上下文配置。
+    - 对支持的聊天模板，默认启用 `preserve_reasoning`；可通过 `--no-reasoning-preserve` 关闭（[#28174](https://github.com/ggml-org/llama.cpp/pull/28174)）。
+    - `input_audio`、`input_video` 请求支持 `data:` URL，修复媒体解码错误（[#27735](https://github.com/ggml-org/llama.cpp/pull/27735)）；需使用兼容的多模态模型与客户端。
+
+    **稳定性与内存**
+    - 降低启用张量重排时模型加载阶段的内存峰值（[#27483](https://github.com/ggml-org/llama.cpp/pull/27483)）。
+    - 修复模型路由器并发请求的 LRU 卡住问题（[#28539](https://github.com/ggml-org/llama.cpp/pull/28539)）、图片输入后的推测解码（[#28715](https://github.com/ggml-org/llama.cpp/pull/28715)）和 CUDA Flash Attention 同步问题（[#27870](https://github.com/ggml-org/llama.cpp/pull/27870)）。
+
+    以上为上游引擎变化，各模型专属能力尚未全部由本次 Chart 升级验证。llm-init 保持 `v1.5.0`。
+    完整上游变更：https://github.com/ggml-org/llama.cpp/compare/b10588...b10969
 
   fullDescription: |
     **重要提示**