[UPDATE][llamacppllmbasev3][1.2.38] Update CUDA engine to b10969

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
+version: 1.2.38

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: '1.2.38'
   categories:
     - AI
     - models
@@ -36,37 +36,30 @@ workloadReplicas:
   llminit: 1
 spec:
   onlyAdmin: true
-  versionName: 'b10588'
+  versionName: 'b10969'
   upgradeDescription: |
-    Chart 1.2.37 — bump engine to `server-cuda12-b10588`.
+    Chart 1.2.38 — Upgrade llama.cpp from `server-cuda12-b10588` 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`.
+    **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.
 
-    134 upstream commits; core libraries move to llama.cpp 0.2.0 / ggml 0.21.0.
+    **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.
 
-    **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))
+    **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.
 
-    **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))
+    **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)).
 
-    **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
+    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
 
-    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.

llamacppllmbasev3/i18n/en-US/OlaresManifest.yaml

@@ -4,25 +4,28 @@ metadata:
 
 spec:
   upgradeDescription: |
-    Chart 1.2.33 — bump engine to `server-cuda12-b10454` and llm-init to `v1.3.8`.
+    Chart 1.2.38 — Upgrade llama.cpp from `server-cuda12-b10588` 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`.
+    **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.
 
-    **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'`.
+    **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.
 
-    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
+    **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.
 
-    **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
+    **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
 
-    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.

llamacppllmbasev3/i18n/zh-CN/OlaresManifest.yaml

@@ -4,25 +4,28 @@ metadata:
 
 spec:
   upgradeDescription: |
-    Chart 1.2.33 — 引擎升级到 `server-cuda12-b10454`，llm-init 升级到 `v1.3.8`。
+    Chart 1.2.38 — llama.cpp 镜像从 `server-cuda12-b10588` 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`。
+    **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。
 
-    **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'`。
+    **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。
 
-    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
+    **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。
 
-    **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
+    **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
 
-    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。

llamacppllmbasev3/templates/llamacpp.yaml

@@ -73,7 +73,7 @@ spec:
             defaultMode: 0555
       containers:
         - name: llamacpp
-          image: docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10588
+          image: docker.io/beclab/ggml-org-llama.cpp:server-cuda12-b10969
           imagePullPolicy: IfNotPresent
           command: ["/bin/sh", "/llm-init/wrappers/llamacpp.sh"]
           env: