From ac274eeef35f3fe6dbde3e0c3b2655b2cda15204 Mon Sep 17 00:00:00 2001 From: Haro Date: Wed, 12 Aug 2026 09:19:11 +0000 Subject: [PATCH] [UPDATE][vllmllmbasev3][1.2.29] --- vllmllmbasev3/Chart.yaml | 4 ++-- vllmllmbasev3/OlaresManifest.yaml | 25 ++++++++++++++++---- vllmllmbasev3/i18n/en-US/OlaresManifest.yaml | 21 ++++++++++++++++ vllmllmbasev3/i18n/zh-CN/OlaresManifest.yaml | 21 ++++++++++++++++ vllmllmbasev3/templates/_helpers.tpl | 6 ++--- vllmllmbasev3/values.yaml | 6 ++--- 6 files changed, 71 insertions(+), 12 deletions(-) diff --git a/vllmllmbasev3/Chart.yaml b/vllmllmbasev3/Chart.yaml index e108d6eb0..21a2e94c4 100644 --- a/vllmllmbasev3/Chart.yaml +++ b/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 diff --git a/vllmllmbasev3/OlaresManifest.yaml b/vllmllmbasev3/OlaresManifest.yaml index 22b357207..a435eb3aa 100644 --- a/vllmllmbasev3/OlaresManifest.yaml +++ b/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. diff --git a/vllmllmbasev3/i18n/en-US/OlaresManifest.yaml b/vllmllmbasev3/i18n/en-US/OlaresManifest.yaml index 83f8a1b4e..55dc73343 100644 --- a/vllmllmbasev3/i18n/en-US/OlaresManifest.yaml +++ b/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. diff --git a/vllmllmbasev3/i18n/zh-CN/OlaresManifest.yaml b/vllmllmbasev3/i18n/zh-CN/OlaresManifest.yaml index 44f3abe10..6930e68b9 100644 --- a/vllmllmbasev3/i18n/zh-CN/OlaresManifest.yaml +++ b/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。 diff --git a/vllmllmbasev3/templates/_helpers.tpl b/vllmllmbasev3/templates/_helpers.tpl index 22e7a76d2..d07a3d9dc 100644 --- a/vllmllmbasev3/templates/_helpers.tpl +++ b/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). */ -}} diff --git a/vllmllmbasev3/values.yaml b/vllmllmbasev3/values.yaml index 35e291fcc..da56bfdcc 100644 --- a/vllmllmbasev3/values.yaml +++ b/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