[NEW][ollamamuseglimmer30bdflashv3][1.0.1]

ollamamuseglimmer30bdflashv3/.helmignore

@@ -0,0 +1,18 @@
+# Patterns to ignore when building packages.
+.DS_Store
+.git/
+.gitignore
+.bzr/
+.bzrignore
+.hg/
+.hgignore
+.svn/
+*.swp
+*.bak
+*.tmp
+*.orig
+*~
+.project
+.idea/
+*.tmproj
+.vscode/

ollamamuseglimmer30bdflashv3/Chart.yaml

@@ -0,0 +1,6 @@
+apiVersion: v2
+appVersion: 0.32.8-rc0
+description: Muse Glimmer 30B Q4_K_M + DFlash served with Ollama + llm-init (GPU only)
+name: ollamamuseglimmer30bdflashv3
+type: application
+version: 1.0.1

ollamamuseglimmer30bdflashv3/OlaresManifest.yaml

@@ -0,0 +1,199 @@
+olaresManifest.version: '0.12.0'
+olaresManifest.type: app
+apiVersion: 'v3'
+metadata:
+  name: ollamamuseglimmer30bdflashv3
+  icon: https://app.cdn.olares.com/appstore/ollama/icon.png
+  description: "Muse Glimmer 30B Q4_K_M + DFlash speculative decoding via Ollama."
+  appid: ollamamuseglimmer30bdflashv3
+  title: Muse Glimmer 30B DFlash
+  version: '1.0.1'
+  categories:
+    - AI
+sharedEntrances:
+  - name: ollamamuseglimmer30bdflashv3
+    host: sharedentrances-api
+    port: 0
+    title: Muse Glimmer 30B DFlash
+    invisible: true
+    authLevel: internal
+    icon: https://app.cdn.olares.com/appstore/ollama/icon.png
+entrances:
+  - name: ollamaclient
+    port: 8090
+    host: download-svc
+    title: Muse Glimmer 30B DFlash
+    icon: https://app.cdn.olares.com/appstore/ollama/icon.png
+    openMethod: window
+    authLevel: internal
+workloadReplicas:
+  ollamamuseglimmer30bdflashv3: 1
+  llminit: 1
+spec:
+  onlyAdmin: true
+  versionName: 'muse-glimmer:30b-q4_K_M-dflash'
+  upgradeDescription: |
+    v1.0.1: fix zh-CN i18n encoding; document DFlash decode throughput (~35-65 tok/s).
+    v1.0.0: initial release - Muse Glimmer 30B Q4_K_M with bundled DFlash drafter via Ollama 0.32.8-rc0.
+  fullDescription: |
+    **Model Overview**
+    Same Muse Glimmer 30B Q4_K_M multimodal agent as the standard tag, plus a bundled **DFlash** speculative-decoding drafter (~1.6GB) for faster generation. Pulls `muse-glimmer:30b-q4_K_M-dflash` (~20GB total, 128K context) from the Ollama library. Output quality matches standard autoregressive decoding; typical decode throughput is **~35-65 tok/s** (GPU- and context-dependent; usually higher than the standard tag on long agent replies).
+    Requires **Ollama 0.32.8-rc0 or newer**.
+    **Preset Configuration**
+    - Model: `muse-glimmer:30b-q4_K_M-dflash`
+    - Capabilities: thinking, tools, vision
+    - Modelfile default: `draft_num_predict=3` (DFlash; set to 0 in a custom Modelfile to disable)
+    - Engine args: `OLLAMA_KEEP_ALIVE=-1 OLLAMA_CONTEXT_LENGTH=131072 OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 OLLAMA_NUM_PARALLEL=1`
+    - Required GPU memory: `22Gi`
+    - Typical decode throughput: **~35-65 tok/s** (GPU-dependent)
+    **Recommended API parameters:** `temperature=1.0`, `top_p=0.95`, `top_k=64`; reasoning effort `high`/`xhigh` for agent tasks.
+    **Accelerator modes**
+    | Mode           | Device            | Details                                                              |
+    |----------------|-------------------|----------------------------------------------------------------------|
+    | `nvidia`       | GPU (CUDA)        | Uses `22Gi` GPU memory                                               |
+    | `nvidia-gb10`  | GPU (Spark)       | Like above, but memory is managed by the pod                         |
+    **Stack**
+    - **Ollama daemon** (port 11434, `ollama/ollama:0.32.8-rc0`)
+    - **Model Console** (port 8090): OpenAI API proxy
+    **Model Storage**
+    Shared at `appCommon/ollama` (blobs dedupe with other Ollama apps).
+  developer: Meta
+  website: https://ollama.com/library/muse-glimmer
+  sourceCode: https://github.com/ollama/ollama
+  submitter: Olares
+  locale:
+    - en-US
+    - zh-CN
+  license:
+    - text: Apache-2.0
+      url: https://www.apache.org/licenses/LICENSE-2.0
+  accelerator:
+    - mode: nvidia
+      limitedCpu: "15360m"
+      requiredCpu: "1300m"
+      requiredDisk: 50Mi
+      limitedDisk: 500Gi
+      limitedMemory: "30Gi"
+      requiredMemory: "20Gi"
+      requiredGPUMemory: "22Gi"
+      limitedGPUMemory: "24Gi"
+    - mode: nvidia-gb10
+      limitedCpu: "15360m"
+      requiredCpu: "1300m"
+      requiredDisk: 50Mi
+      limitedDisk: 500Gi
+      limitedMemory: "28Gi"
+      requiredMemory: "24Gi"
+  supportArch:
+    - amd64
+    - arm64
+permission:
+  appData: true
+  appCommon: true
+options:
+  apiTimeout: 0
+  LLMGatewaySupported: true
+  shared: true
+  templateOnly: false
+  dependencies:
+    - name: olares
+      version: '>=1.12.6-0'
+      type: system
+envs:
+  - envName: MODEL_SOURCE
+    required: true
+    type: string
+    editable: false
+    applyOnChange: false
+    default: "ollama://muse-glimmer:30b-q4_K_M-dflash"
+    description: "Fixed to ollama://muse-glimmer:30b-q4_K_M-dflash (main model + DFlash drafter)."
+    regex: "^ollama://"
+  - envName: MODEL_NAME
+    required: true
+    type: string
+    editable: false
+    applyOnChange: false
+    default: "muse-glimmer:30b-q4_K_M-dflash"
+    description: "OpenAI alias. Fixed to muse-glimmer:30b-q4_K_M-dflash."
+  - envName: MODEL_MODE
+    required: true
+    type: string
+    editable: false
+    applyOnChange: false
+    default: "chat"
+    options:
+    - title: Chat
+      value: chat
+    - title: Embedding
+      value: embedding
+  - envName: MODEL_SUPPORTS
+    required: true
+    type: string
+    editable: false
+    applyOnChange: false
+    default: "thinking,tools,vision"
+    multiSelect: true
+    splitter: ","
+    options:
+    - title: Vision
+      value: vision
+    - title: Tools
+      value: tools
+    - title: Thinking
+      value: thinking
+    - title: None
+      value: none
+  - envName: ENGINE_ARGS
+    required: true
+    default: "OLLAMA_KEEP_ALIVE=-1 OLLAMA_CONTEXT_LENGTH=131072 OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 OLLAMA_NUM_PARALLEL=1"
+    type: string
+    editable: true
+    applyOnChange: true
+    description: "Ollama daemon env. DFlash draft count is in the Ollama Modelfile (draft_num_predict=3), not here."
+  - envName: LOG_LEVEL
+    required: false
+    type: string
+    editable: true
+    applyOnChange: true
+    default: "info"
+    options:
+    - title: Debug
+      value: debug
+    - title: Info
+      value: info
+    - title: Warn
+      value: warn
+    - title: Error
+      value: error
+  - envName: OLLAMA_CPU_REQUEST
+    required: true
+    default: "200m"
+    type: string
+    editable: false
+    applyOnChange: false
+  - envName: OLLAMA_MEMORY_REQUEST
+    required: true
+    default: "19Gi"
+    type: string
+    editable: false
+    applyOnChange: false
+  - envName: OLLAMA_CPU_LIMIT
+    required: true
+    default: "6"
+    type: string
+    editable: false
+    applyOnChange: false
+  - envName: OLLAMA_MEMORY_LIMIT
+    required: true
+    default: "28Gi"
+    type: string
+    editable: false
+    applyOnChange: false
+  - envName: OLLAMA_REQUIRED_GPU_MEMORY
+    required: true
+    type: string
+    editable: false
+    applyOnChange: false
+    default: "22Gi"
+    description: "GPU memory for ~20GB tag (main + DFlash drafter) plus 128K KV headroom."
+    regex: "^[0-9]+(Gi|Mi)?$"

ollamamuseglimmer30bdflashv3/i18n/en-US/OlaresManifest.yaml

@@ -0,0 +1,24 @@
+metadata:
+
+  description: "Muse Glimmer 30B Q4_K_M + DFlash speculative decoding via Ollama."
+
+  title: Muse Glimmer 30B DFlash
+
+
+
+spec:
+
+  fullDescription: |
+
+    **Model Overview**
+
+    Same Muse Glimmer 30B Q4_K_M as the standard tag, with bundled DFlash drafter for faster generation. Tag: `muse-glimmer:30b-q4_K_M-dflash` (~20GB, 128K context). Requires Ollama 0.32.8-rc0+.
+
+
+
+    **Preset:** thinking + tools + vision; `22Gi` GPU; Modelfile `draft_num_predict=3`; typical decode throughput **~35-65 tok/s** (GPU-dependent).
+
+
+
+    **Stack:** Ollama daemon (11434) + Model Console (8090). Models stored at `appCommon/ollama`.
+

ollamamuseglimmer30bdflashv3/i18n/zh-CN/OlaresManifest.yaml

@@ -0,0 +1,33 @@
+metadata:
+  description: "Muse Glimmer 30B Q4_K_M + DFlash 推测解码加速，基于 Ollama 运行。"
+  title: Muse Glimmer 30B DFlash
+
+spec:
+  fullDescription: |
+    **模型概述**
+    与标准 `30b-q4_K_M` 标签相同的主模型，额外 bundled DFlash drafter（约 1.6GB）用于更快生成。从 Ollama 库拉取 `muse-glimmer:30b-q4_K_M-dflash`（约 20GB，128K 上下文）。输出质量与标准自回归解码一致；长 Agent 回复场景吞吐通常更高。
+
+    需要 **Ollama 0.32.8-rc0 或更新版本**。
+
+    **预设配置**
+    - 模型：`muse-glimmer:30b-q4_K_M-dflash`
+    - 能力：thinking、tools、vision
+    - Modelfile 默认：`draft_num_predict=3`（DFlash；自定义 Modelfile 设为 0 可关闭）
+    - 引擎参数：`OLLAMA_KEEP_ALIVE=-1 OLLAMA_CONTEXT_LENGTH=131072 OLLAMA_FLASH_ATTENTION=1 OLLAMA_KV_CACHE_TYPE=q8_0 OLLAMA_NUM_PARALLEL=1`
+    - 所需 GPU 显存：`22Gi`
+    - 典型解码吞吐：约 **35–65 tok/s**（视 GPU 与上下文长度而定；通常高于标准标签）
+
+    **推荐 API 参数：** `temperature=1.0`、`top_p=0.95`、`top_k=64`；复杂 Agent 任务使用较高 reasoning effort（`high`/`xhigh`）。
+
+    **加速器模式**
+    | 模式           | 设备             | 说明                                                         |
+    |----------------|------------------|--------------------------------------------------------------|
+    | `nvidia`       | GPU (CUDA)       | 使用 `22Gi` 显存                                             |
+    | `nvidia-gb10`  | GPU (Spark)      | 类似上方，显存由 pod 管理                                    |
+
+    **技术栈**
+    - **Ollama 守护进程**（11434 端口）
+    - **Model Console**（8090 端口）：OpenAI API 代理
+
+    **模型存储**
+    位于共享目录 `appCommon/ollama`，与其他 Ollama 应用共用（blob 按 digest 去重）。

ollamamuseglimmer30bdflashv3/owners

@@ -0,0 +1,7 @@
+owners:
+- 'LittleLollipop'
+- 'TShentu'
+- 'hysyeah'
+- 'pengpeng'
+- 'harveyff'
+- 'FantasticCode2019'

ollamamuseglimmer30bdflashv3/templates/_helpers.tpl

@@ -0,0 +1,47 @@
+{{- /* llmbase.gpuMiB: normalize a GPU-memory quantity to a BARE MiB integer for
+       HAMi's nvidia.com/gpumem. Its base unit is MiB and the value MUST be a
+       plain integer — a Mi/Gi suffix is misread by the scheduler (e.g. "6144Mi"
+       -> 6442450944). Accepts 8Gi / 8G / 8192Mi / 8192M / 8192 and returns MiB.
+       Usage: {{ include "llmbase.gpuMiB" ($oe.X_REQUIRED_GPU_MEMORY | default "4096") }} */ -}}
+{{- define "llmbase.gpuMiB" -}}
+{{- $g := trim . -}}
+{{- if hasSuffix "Gi" $g -}}
+{{- mul (int (trimSuffix "Gi" $g)) 1024 -}}
+{{- else if hasSuffix "G" $g -}}
+{{- mul (int (trimSuffix "G" $g)) 1024 -}}
+{{- else if hasSuffix "Mi" $g -}}
+{{- int (trimSuffix "Mi" $g) -}}
+{{- else if hasSuffix "M" $g -}}
+{{- int (trimSuffix "M" $g) -}}
+{{- else -}}
+{{- int $g -}}
+{{- end -}}
+{{- end -}}
+{{- /* ollamamuseglimmer30bdflashv3.engineArgs: merge clone ENGINE_ARGS with chart defaults.
+       OLLAMA_KEEP_ALIVE=-1 (forever in VRAM) on the ollama daemon unless set.
+       llm-init ENGINE_ARGS strips this key (see engineArgsForLlminit).
+       Usage: {{ include "ollamamuseglimmer30bdflashv3.engineArgs" (dict "Args" ($oe.ENGINE_ARGS | default "")) }} */ -}}
+{{- define "ollamamuseglimmer30bdflashv3.engineArgs" -}}
+{{- $in := . -}}
+{{- $args := trim ($in.Args | default "") -}}
+{{- if not (contains "OLLAMA_KEEP_ALIVE" $args) -}}
+{{- if $args -}}
+{{- $args = printf "%s OLLAMA_KEEP_ALIVE=-1" $args -}}
+{{- else -}}
+{{- $args = "OLLAMA_KEEP_ALIVE=-1" -}}
+{{- end -}}
+{{- end -}}
+{{- $args -}}
+{{- end -}}
+{{- /* ollamamuseglimmer30bdflashv3.engineArgsForLlminit: user ENGINE_ARGS for llm-init only.
+       Never inject OLLAMA_KEEP_ALIVE (llm-init forwards it on /api/chat as "-1").
+       Strip if user set it.
+       Usage: {{ include "ollamamuseglimmer30bdflashv3.engineArgsForLlminit" (dict "Args" ($oe.ENGINE_ARGS | default "")) }} */ -}}
+{{- define "ollamamuseglimmer30bdflashv3.engineArgsForLlminit" -}}
+{{- $in := . -}}
+{{- $args := trim ($in.Args | default "") -}}
+{{- $args = regexReplaceAll ` ?OLLAMA_KEEP_ALIVE=[^ ]+` "" $args -}}
+{{- $args = regexReplaceAll ` +` " " $args -}}
+{{- $args = trim $args -}}
+{{- $args -}}
+{{- end -}}