[UPDATE][vllmllmbasev3][1.2.29]

vllmllmbasev3/Chart.yaml

@@ -1,6 +1,6 @@
 apiVersion: v2
-appVersion: 0.25.0
+appVersion: 0.27.1
 description: Generic vLLM + llm-init base; the model is supplied at install time via env
 name: vllmllmbasev3
 type: application
-version: 1.2.27
+version: 1.2.29

vllmllmbasev3/OlaresManifest.yaml

@@ -7,7 +7,7 @@ metadata:
   description: "Generic vLLM engine base. Pick any HuggingFace Safetensors model at install via env."
   appid: vllmllmbasev3
   title: vLLM Engine Base
-  version: '1.2.27'
+  version: '1.2.29'
   categories:
     - AI
 sharedEntrances:
@@ -33,11 +33,28 @@ workloadReplicas:
   llminit: 1
 spec:
   onlyAdmin: true
-  versionName: '0.25.0'
+  versionName: '0.27.1'
   upgradeDescription: |
-    v1.2.27: Move `HF_TOKEN` into Kubernetes Secret (`secretKeyRef`); map HF env from `OLARES_SYSTEM_HUGGINGFACE_*`.
+    Upgrade vLLM from v0.26.0 to v0.27.1. Chart 1.2.29.
 
-    HF_ENDPOINT / HF_TOKEN from OLARES_SYSTEM_HUGGINGFACE_* only. Chart 1.2.26.
+    **What's Changed** (relevant to Olares)
+
+    v0.27.1
+    - Support quantized DSpark Markov heads
+
+    v0.27.0
+    - Kimi K3 full-stack support (kernels, DeepGEMM, compressed-tensors NVFP4, DSpark AR fusion)
+    - New models: Qwen3.5 dense/MoE, K-EXAONE-2.0, VaultGemma, jina-embeddings-v5; Inkling NVFP4/FP8
+    - PyTorch 2.13.0 upgrade (breaking environment change)
+    - FlashAttention 4 on SM100: FP8 KV cache and headdim-256; JIT warmup removes first-request stalls
+    - DeepSeek-V4: sequence parallelism, router/topk optimizations, compact MXFP4 indexer KV cache
+    - Model Runner V2: encoder-only attention, embedding/classification pooling, multimodal on CPU, multi-layer MTP
+    - KV offloading: tiered storage, P2P secondary tier, pluggable eviction policies
+    - Large-scale serving: fault-tolerance framework for DP+EP; NIXL P/D for hybrid MLA+SSM models
+    - Rust frontend: gRPC control plane (health, abort, model/KV discovery); `vllm-bench` in CLI
+
+    For detailed release notes, visit:
+    https://github.com/vllm-project/vllm/releases
   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.

vllmllmbasev3/i18n/en-US/OlaresManifest.yaml

@@ -3,6 +3,27 @@ metadata:
   title: vLLM Engine Base
 
 spec:
+  upgradeDescription: |
+    Upgrade vLLM from v0.26.0 to v0.27.1. Chart 1.2.29.
+
+    **What's Changed** (relevant to Olares)
+
+    v0.27.1
+    - Support quantized DSpark Markov heads
+
+    v0.27.0
+    - Kimi K3 full-stack support (kernels, DeepGEMM, compressed-tensors NVFP4, DSpark AR fusion)
+    - New models: Qwen3.5 dense/MoE, K-EXAONE-2.0, VaultGemma, jina-embeddings-v5; Inkling NVFP4/FP8
+    - PyTorch 2.13.0 upgrade (breaking environment change)
+    - FlashAttention 4 on SM100: FP8 KV cache and headdim-256; JIT warmup removes first-request stalls
+    - DeepSeek-V4: sequence parallelism, router/topk optimizations, compact MXFP4 indexer KV cache
+    - Model Runner V2: encoder-only attention, embedding/classification pooling, multimodal on CPU, multi-layer MTP
+    - KV offloading: tiered storage, P2P secondary tier, pluggable eviction policies
+    - Large-scale serving: fault-tolerance framework for DP+EP; NIXL P/D for hybrid MLA+SSM models
+    - Rust frontend: gRPC control plane (health, abort, model/KV discovery); `vllm-bench` in CLI
+
+    For detailed release notes, visit:
+    https://github.com/vllm-project/vllm/releases
   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.

vllmllmbasev3/i18n/zh-CN/OlaresManifest.yaml

@@ -3,6 +3,27 @@ metadata:
   title: vLLM 引擎基座
 
 spec:
+  upgradeDescription: |
+    将 vLLM 从 v0.26.0 升级至 v0.27.1。Chart 1.2.29。
+
+    **更新内容**（与 Olares 相关）
+
+    v0.27.1
+    - 支持量化 DSpark Markov heads
+
+    v0.27.0
+    - Kimi K3 全栈支持（内核、DeepGEMM、compressed-tensors NVFP4、DSpark AR fusion）
+    - 新增模型：Qwen3.5 dense/MoE、K-EXAONE-2.0、VaultGemma、jina-embeddings-v5；Inkling NVFP4/FP8
+    - PyTorch 2.13.0 升级（破坏性环境变更）
+    - SM100 上 FlashAttention 4：FP8 KV cache 与 headdim-256；JIT warmup 消除首请求编译卡顿
+    - DeepSeek-V4：序列并行、router/topk 优化、紧凑 MXFP4 indexer KV cache
+    - Model Runner V2：encoder-only 注意力、embedding/分类 pooling、CPU 多模态、多层 MTP
+    - KV offloading：分层存储、P2P 二级 tier、可插拔淘汰策略
+    - 大规模 serving：DP+EP 容错框架；混合 MLA+SSM 模型的 NIXL P/D
+    - Rust 前端：gRPC 控制面（健康检查、abort、模型/KV 发现）；CLI 集成 `vllm-bench`
+
+    详细发布说明请访问：
+    https://github.com/vllm-project/vllm/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。

vllmllmbasev3/templates/_helpers.tpl

@@ -57,11 +57,11 @@ amd64
 {{- $arch := include "llmbase.hostArch" . -}}
 {{- $img := .Values.engine.images | default dict -}}
 {{- if $isGb10 -}}
-{{- $img.nvidiaGb10 | default "docker.io/vllm/vllm-openai:latest-aarch64-cu130" -}}
+{{- $img.nvidiaGb10 | default "docker.io/vllm/vllm-openai:v0.27.1-aarch64" -}}
 {{- else if eq $arch "arm64" -}}
-{{- $img.nvidiaArm64 | default "docker.io/vllm/vllm-openai:v0.24.0-aarch64-cu129" -}}
+{{- $img.nvidiaArm64 | default "docker.io/vllm/vllm-openai:v0.27.1-aarch64" -}}
 {{- else -}}
-{{- $img.nvidia | default "docker.io/vllm/vllm-openai:v0.25.0" -}}
+{{- $img.nvidia | default "docker.io/vllm/vllm-openai:v0.27.1" -}}
 {{- end -}}
 {{- end -}}
 {{- /* Spark/GB10: detect from GPU.Type or node hardware (install-time .Values.nodes). */ -}}

vllmllmbasev3/values.yaml

@@ -11,6 +11,6 @@ workloads:
 # vLLM engine images by Olares accelerator mode (override in fork/preset if needed).
 engine:
   images:
-    nvidia: docker.io/vllm/vllm-openai:v0.25.0
-    nvidiaArm64: docker.io/vllm/vllm-openai:v0.25.0-aarch64
-    nvidiaGb10: docker.io/vllm/vllm-openai:v0.25.0-aarch64
+    nvidia: docker.io/vllm/vllm-openai:v0.27.1
+    nvidiaArm64: docker.io/vllm/vllm-openai:v0.27.1-aarch64
+    nvidiaGb10: docker.io/vllm/vllm-openai:v0.27.1-aarch64